MnemoCore

Infrastructure for Persistent Cognitive Memory

"Memory is not a container. It is a living process — a holographic continuum where every fragment contains the whole."

Quick Install

Install v3.0.0 after publication:

pip install mnemocore==3.0.0

Install from source (development):

git clone https://github.com/RobinALG87/MnemoCore-Persistent-Cognitive-Ai-Memory.git
cd MnemoCore-Persistent-Cognitive-Ai-Memory
python -m venv .venv && .\.venv\Scripts\activate   # Windows
# source .venv/bin/activate                        # Linux / macOS
pip install -e .          # editable install
pip install -e ".[dev]"   # + pytest, mypy, black, etc.

For the REST server, set your API key before starting:
# Windows PowerShell
$env:HAIM_API_KEY = "your-secure-key"
# Linux / macOS
# export HAIM_API_KEY="your-secure-key"
Then start the legacy v2 API locally: mnemocore.

mnemocore starts the legacy v2 REST server; mnemocore-cli is the legacy Click CLI. Neither is the v3 AgentMemory persistence boundary. For v3, compose HybridMemoryRuntime over an explicit scope or deploy create_v3_app with a credential-to-scope authorizer. AgentMemory itself is local-first and does not require a running Redis or Qdrant service. Full legacy setup, Docker, and configuration details are in Installation below.

MnemoCore 3.0.0 beta

MnemoCore is persistent memory and context infrastructure for agents. Version 3.0.0 beta releases an AgentMemory-first runtime: local-first, exact-scope SQLite memory with timeline/history support, deterministic context compilation, and rebuildable projections. HybridMemoryRuntime adds deterministic lexical and BinaryHDV retrieval without adding a second persistence path.

The v3 contract is explicit:

Every operation requires a complete MemoryScope; cross-scope fallback is forbidden.
AgentMemory is the only durable source of truth for v3.
Cognitive effects are validated plans, not autonomous direct writes.
The v3 HTTP application is composed with a scope authorizer and fails closed when one is absent.
v2 HAIM, REST, MCP, and CLI paths remain compatibility-only and are not the v3 persistence layer.

Version 3.0 retains the established dependency set for installation compatibility. A dependency-minimal AgentMemory distribution remains a future major-release migration.

Legacy cognitive capabilities (v2 compatibility only)

The following HAIM description is retained as a reference for existing v2 deployments. It is not a v3 persistence or deployment contract.

Traditional vector stores retrieve. MnemoCore thinks. It is built on the mathematical framework of Binary Hyperdimensional Computing (HDC) and Vector Symbolic Architectures (VSA), principles rooted in Pentti Kanerva's landmark 2009 theory of cognitive computing. Every memory is encoded as a 16,384-dimensional binary holographic vector — a format that is simultaneously compact (2,048 bytes), noise-tolerant (Hamming geometry), and algebraically rich (XOR binding, majority bundling, circular permutation).

At its core lives the Holographic Active Inference Memory (HAIM) Engine — a system that does not merely answer queries, but:

Evaluates the epistemic novelty of every incoming memory before deciding to store it
Dreams — strengthening synaptic connections between related memories during idle cycles
Reasons by analogy — if king:man :: ?:woman, the VSA soul computes queen
Self-organizes into tiered storage based on biologically-inspired Long-Term Potentiation (LTP)
Scales from a single process to distributed nodes targeting 1B+ memories

Phase 4.x introduces cognitive enhancements including contextual masking, reliability feedback loops, semantic consolidation, gap detection/filling, temporal recall (episodic chaining + chrono-weighted query), a Subconscious Daemon with LLM-powered dream synthesis, and a full dependency-injection container pattern for clean modularity.

Phase 5 transforms MnemoCore into a true cognitive architecture with Working Memory, Episodic/Semantic/Procedural stores, Meta Memory, and a Self-Improvement Worker — all orchestrated by an 11-phase Pulse loop.

Phase 6 adds four research-backed services: StrategyBank (reasoning bank with Bayesian confidence), KnowledgeGraph (spreading activation + community detection), MemoryScheduler (OS-level priority queue with interrupts), and SAMEP (multi-agent memory exchange with HMAC integrity).

Architecture
Core Technology
The Memory Lifecycle
Tiered Storage
Phase 4.0 Cognitive Enhancements
Phase 4.4–4.5 Subconscious Daemon & LLM Integration
API Reference
Python Library Usage
Installation
Configuration
MCP Server Integration
Observability
Roadmap
Contributing

Architecture

The diagram below documents the legacy v2 architecture. New v3 deployments use the scoped AgentMemory and HybridMemoryRuntime composition described above.

┌─────────────────────────────────────────────────────────────────┐
│                        MnemoCore Stack                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│   ┌──────────────────────────────────────────────────────────┐  │
│   │              REST API  (FastAPI / Async)                  │  │
│   │   /store  /query  /feedback  /insights/gaps  /stats      │  │
│   │   Rate Limiting · API Key Auth · Prometheus Metrics      │  │
│   └─────────────────────────┬────────────────────────────────┘  │
│                             │                                   │
│   ┌─────────────────────────▼────────────────────────────────┐  │
│   │                  HAIM Engine                              │  │
│   │                                                          │  │
│   │   ┌──────────────┐  ┌──────────────┐  ┌──────────────┐  │  │
│   │   │ Text Encoder │  │  EIG / Epist │  │  Subconsc.   │  │  │
│   │   │ (token→HDV)  │  │  Drive       │  │  Dream Loop  │  │  │
│   │   └──────────────┘  └──────────────┘  └──────────────┘  │  │
│   │                                                          │  │
│   │   ┌──────────────────────────────────────────────────┐  │  │
│   │   │            Binary HDV Core (VSA)                 │  │  │
│   │   │  XOR bind · majority_bundle · permute · Hamming  │  │  │
│   │   └──────────────────────────────────────────────────┘  │  │
│   └─────────────────────────┬────────────────────────────────┘  │
│                             │                                   │
│   ┌─────────────────────────▼────────────────────────────────┐  │
│   │                  Tier Manager                             │  │
│   │                                                          │  │
│   │   🔥 HOT         🌡 WARM            ❄️ COLD               │  │
│   │   In-Memory      Redis / mmap       Qdrant / Disk / S3   │  │
│   │   ≤2,000 nodes   ≤100,000 nodes     ∞ nodes              │  │
│   │   <1ms           <10ms              <100ms               │  │
│   └──────────────────────────────────────────────────────────┘  │
│                                                                 │
│   ┌─────────────────────────────────────────────────────────┐   │
│   │                  Conceptual Layer ("The Soul")           │   │
│   │   ConceptualMemory · Analogy Engine · Symbol Algebra     │   │
│   └─────────────────────────────────────────────────────────┘   │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Component Overview

Component	File	Responsibility
HAIM Engine	`src/mnemocore/core/engine.py`	Central cognitive coordinator — store, query, dream, delete
BinaryHDV	`src/mnemocore/core/binary_hdv.py`	16384-dim binary vector math (XOR, Hamming, bundle, permute)
TextEncoder	`src/mnemocore/core/binary_hdv.py`	Token→HDV pipeline with positional permutation binding
MemoryNode	`src/mnemocore/core/node.py`	Memory unit with LTP, epistemic values, tier state
TierManager	`src/mnemocore/core/tier_manager.py`	HOT/WARM/COLD orchestration with LTP-driven eviction
SynapticConnection	`src/mnemocore/core/synapse.py`	Hebbian synapse with strength, decay, and fire tracking
SynapseIndex	`src/mnemocore/core/synapse_index.py`	Fast synapse lookup index for associative spreading
ConceptualMemory	`src/mnemocore/core/holographic.py`	VSA soul for analogy and cross-domain symbolic reasoning
AsyncRedisStorage	`src/mnemocore/core/async_storage.py`	Async Redis backend (WARM tier + pub/sub)
BayesianLTP	`src/mnemocore/core/bayesian_ltp.py`	Bayesian reliability scoring on top of LTP strength
SemanticConsolidation	`src/mnemocore/core/semantic_consolidation.py`	Memory deduplication via majority-bundle prototyping
ConsolidationWorker	`src/mnemocore/core/consolidation_worker.py`	Async worker scheduling nightly consolidation
GapDetector	`src/mnemocore/core/gap_detector.py`	Temporal co-occurrence analysis for knowledge gaps
GapFiller	`src/mnemocore/core/gap_filler.py`	Bridge detected gaps via synapse creation
Immunology	`src/mnemocore/core/immunology.py`	Auto-associative attractor cleanup for vector drift
Attention	`src/mnemocore/core/attention.py`	XOR context masking / project isolation
BatchOps	`src/mnemocore/core/batch_ops.py`	Vectorized bulk store / query operations
HNSWIndex	`src/mnemocore/core/hnsw_index.py`	In-process HNSW approximate nearest-neighbour index
QdrantStore	`src/mnemocore/core/qdrant_store.py`	Async Qdrant COLD tier backend
RecursiveSynthesizer	`src/mnemocore/core/recursive_synthesizer.py`	Deep concept synthesis via iterative VSA composition
RippleContext	`src/mnemocore/core/ripple_context.py`	Cascading context propagation across synaptic graph
SubconsciousAI	`src/mnemocore/core/subconscious_ai.py`	LLM-guided dream synthesis worker
SubconsciousDaemon	`src/mnemocore/subconscious/daemon.py`	Background process orchestrating dream/consolidation cycles
LLMIntegration	`src/mnemocore/llm_integration.py`	Agent-facing LLM connector (OpenAI / Anthropic compatible)
Container	`src/mnemocore/core/container.py`	Dependency-injection wiring for all core components
GoalTree	`src/mnemocore/meta/goal_tree.py`	Hierarchical goal / task tracking for meta-cognition
LearningJournal	`src/mnemocore/meta/learning_journal.py`	Persistent log of what the agent has learned over time
API	`src/mnemocore/api/main.py`	Legacy v2 HAIM REST interface; v3 uses `create_v3_app` with scope authorization
MCP Server	`src/mnemocore/mcp/server.py`	Legacy v2 HAIM REST adapter; v3 uses scoped AgentMemory integration helpers

Core Technology: Binary HDV & VSA

MnemoCore's mathematical foundation is Hyperdimensional Computing — a computing paradigm that encodes information in very high-dimensional binary vectors (HDVs), enabling noise-tolerant, distributed, and algebraically composable representations.

The Vector Space

Every piece of information — a word, a sentence, a concept, a goal — is encoded as a 16,384-dimensional binary vector:

Dimension D = 16,384 bits = 2,048 bytes per vector
Storage:      packed as numpy uint8 arrays
Similarity:   Hamming distance (popcount of XOR result)
Random pair:  ~50% similarity (orthogonality by probability)

At this dimensionality, two random vectors will differ in ~50% of bits. This near-orthogonality is the foundation of the system's expressive power — related concepts cluster together while unrelated ones remain maximally distant.

VSA Algebra

Four primitive operations make the entire system work:

Binding — XOR `⊕`

Creates an association between two concepts. Crucially, the result is dissimilar to both inputs (appears as noise), making it a true compositional operation.

# Bind content to its context
bound = content_vec.xor_bind(context_vec)  # content ⊕ context

# Self-inverse: unbind by re-binding
recovered = bound.xor_bind(context_vec)   # ≈ content (XOR cancels)

Key mathematical properties:

Self-inverse: A ⊕ A = 0 (XOR cancels itself)
Commutative: A ⊕ B = B ⊕ A
Distance-preserving: hamming(A⊕C, B⊕C) = hamming(A, B)

Bundling — Majority Vote

Creates a prototype that is similar to all inputs. This is how multiple memories combine into a concept.

from mnemocore.core.binary_hdv import majority_bundle

# Create semantic prototype from related memories
concept = majority_bundle([vec_a, vec_b, vec_c, vec_d])  # similar to all inputs

Permutation — Circular Shift

Encodes sequence and roles without separate positional embeddings.

# Positional encoding: token at position i
positioned = token_vec.permute(shift=i)  # circular bit-shift

# Encode "hello world" with order information
hello_positioned = encoder.get_token_vector("hello").permute(0)
world_positioned = encoder.get_token_vector("world").permute(1)
sentence_vec = majority_bundle([hello_positioned, world_positioned])

Similarity — Hamming Distance

Fast comparison using vectorized popcount over XOR results:

# Normalized similarity: 1.0 = identical, 0.5 = unrelated
sim = vec_a.similarity(vec_b)  # 1.0 - hamming(a, b) / D

# Batch nearest-neighbor search (no Python loops)
distances = batch_hamming_distance(query, database_matrix)

Text Encoding Pipeline

The TextEncoder converts natural language to HDVs using a token-position binding scheme:

"Python TypeError" →
  token_hdv("python") ⊕ permute(0)  =  positioned_0
  token_hdv("typeerror") ⊕ permute(1)  =  positioned_1
  majority_bundle([positioned_0, positioned_1])  =  final_hdv

Token vectors are deterministic — seeded via SHAKE-256 hash — meaning the same word always produces the same base vector, enabling cross-session consistency without a vocabulary file.

The Memory Lifecycle

Every memory passes through a defined lifecycle from ingestion to long-term storage:

Incoming Content
      │
      ▼
 ┌─────────────┐
 │ TextEncoder │  → 16,384-dim binary HDV
 └──────┬──────┘
        │
        ▼
 ┌──────────────────┐
 │ Context Binding  │  → XOR bind with goal_context if present
 │  (XOR)           │     bound_vec = content ⊕ context
 └──────┬───────────┘
        │
        ▼
 ┌──────────────────┐
 │  EIG Evaluation  │  → Epistemic Information Gain
 │  (Novelty Check) │     eig = normalized_distance(vec, context_vec)
 └──────┬───────────┘     tag "epistemic_high" if eig > threshold
        │
        ▼
 ┌─────────────────┐
 │  MemoryNode     │  → id, hdv, content, metadata
 │  Creation       │     ltp_strength = I × log(1+A) × e^(-λT)
 └──────┬──────────┘
        │
        ▼
 ┌─────────────────┐
 │  HOT Tier       │  → In-memory dict (max 2000 nodes)
 │  (RAM)          │     LTP eviction: low-LTP nodes → WARM
 └──────┬──────────┘
        │     (background)
        ▼
 ┌─────────────────┐
 │ Subconscious    │  → Dream cycle fires
 │ Dream Loop      │     Query similar memories
 └──────┬──────────┘     Strengthen synapses (Hebbian)
        │
        ▼
 ┌─────────────────┐
 │  WARM Tier      │  → Redis-backed persistence
 │  (Redis/mmap)   │     async dual-write + pub/sub events
 └──────┬──────────┘
        │     (scheduled, nightly)
        ▼
 ┌─────────────────┐
 │  COLD Tier      │  → Qdrant / Disk / S3
 │  (Archival)     │     ANN search, long-term persistence
 └─────────────────┘

Long-Term Potentiation (LTP)

Memories are not equal. Importance is computed dynamically using a biologically-inspired LTP formula:

S = I × log(1 + A) × e^(-λ × T)

Where:
  S = LTP strength (determines tier placement)
  I = Importance (derived from epistemic + pragmatic value)
  A = Access count (frequency of retrieval)
  λ = Decay lambda (configurable, default ~0.01)
  T = Age in days

Memories with high LTP remain in HOT tier. Those that decay are automatically promoted to WARM, then COLD — mirroring how biological memory consolidates from working memory to long-term storage.

Synaptic Connections

Memories are linked by SynapticConnection objects that implement Hebbian learning: "neurons that fire together, wire together."

Every time two memories are co-retrieved (via the background dream loop or explicit binding), their synaptic strength increases. During query time, synaptic spreading amplifies scores of connected memories even when they do not directly match the query vector — enabling associative recall.

# Explicit synapse creation
engine.bind_memories(id_a, id_b, success=True)

# Associative spreading: query top seeds spread activation to neighbors
# neighbor_score += seed_score × synapse_strength × 0.3

Tiered Storage: HOT / WARM / COLD

Tier	Backend	Capacity	Latency	Eviction Trigger
🔥 HOT	Python dict (RAM)	2,000 nodes	< 1ms	LTP < threshold
🌡 WARM	Redis + mmap	100,000 nodes	< 10ms	Age + low access
❄️ COLD	Qdrant / Disk / S3	Unlimited	< 100ms	Manual / scheduled

Promotion is automatic: accessing a WARM or COLD memory re-promotes it to HOT based on recalculated LTP. Eviction is LRU-weighted by LTP strength — the most biologically active memories always stay hot.

Phase 4.0 Cognitive Enhancements

MnemoCore Phase 4.0 introduces five architectural enhancements that elevate the system from data retrieval to cognitive reasoning.

1. Contextual Query Masking (XOR Attention)

Problem: Large multi-project deployments suffer from cross-context interference. A query for "Python error handling" returns memories from all projects equally, diluting precision.

Solution: Bidirectional XOR context binding — apply the same context vector at both storage and query time:

Store:  bound_vec   = content ⊕ context_vec
Query:  masked_query = query   ⊕ context_vec

Result: (content ⊕ C) · (query ⊕ C) ≈ content · query
        (context cancels, cross-project noise is suppressed)

# Store memories in a project context
engine.store("API rate limiting logic", goal_id="ProjectAlpha")
engine.store("Garden watering schedule", goal_id="HomeProject")

# Query with context mask — only ProjectAlpha memories surface
results = engine.query("API logic", top_k=5, context="ProjectAlpha")

Expected impact: +50–80% query precision (P@5) in multi-project deployments.

2. Reliability Feedback Loop (Self-Correcting Memory)

Problem: Wrong or outdated memories persist with the same retrieval weight as correct ones. The system has no mechanism to learn from its own mistakes.

Solution: Bayesian reliability scoring with real-world outcome feedback:

reliability = (successes + 1) / (successes + failures + 2)  # Laplace smoothing

LTP_enhanced = I × log(1+A) × e^(-λT) × reliability

# After using a retrieved memory:
engine.provide_feedback(memory_id, outcome=True)   # Worked → boost reliability
engine.provide_feedback(memory_id, outcome=False)  # Failed → reduce reliability

# System auto-tags consistently wrong memories as "unreliable"
# and verified memories (>5 successes, >0.8 score) as "verified"

The system converges toward high-confidence knowledge — memories that have demonstrably worked in practice rank above theoretically similar but unproven ones.

3. Semantic Memory Consolidation (Dream-Phase Synthesis)

Problem: Episodic memory grows without bound. 1,000 memories about "Python TypeError" are semantically equivalent but consume 2MB of vector space and slow down linear scan queries.

Solution: Nightly ConsolidationWorker clusters similar WARM tier memories and replaces them with a semantic anchor — a majority-bundled prototype:

BEFORE consolidation:
  mem_001: "Python TypeError in line 45"    (2KB vector)
  mem_002: "TypeError calling function"     (2KB vector)
  ...   ×100 similar memories              (200KB total)

AFTER consolidation:
  anchor_001: "Semantic pattern: python typeerror function"
              metadata: {source_count: 100, confidence: 0.94}
              hdv: majority_bundle([mem_001.hdv, ..., mem_100.hdv])  (2KB)

# Manual trigger (runs automatically at 3 AM)
stats = engine.trigger_consolidation()
# → {"abstractions_created": 12, "memories_consolidated": 847}

# Via API (admin endpoint)
POST /admin/consolidate

Expected impact: 70–90% memory footprint reduction, 10x query speedup at scale.

4. Auto-Associative Cleanup Loop (Vector Immunology)

Problem: Holographic vectors degrade over time through repeated XOR operations, noise accumulation, and long-term storage drift. After months of operation, retrieved vectors become "blurry" and similarity scores fall.

Solution: Iterative attractor dynamics — when a retrieved vector appears noisy, snap it to the nearest stable concept in a codebook of high-confidence prototypes:

noisy_vec → find K nearest in codebook
         → majority_bundle(K neighbors)
         → check convergence (Hamming distance < 5%)
         → iterate until converged or max iterations reached

# Cleanup runs automatically on retrieval when noise > 15%
node = engine.get_memory(memory_id, auto_cleanup=True)
# node.metadata["cleaned"] = True  (if cleanup was triggered)
# node.metadata["cleanup_iterations"] = 3

# Codebook is auto-populated from most-accessed, high-reliability memories

Expected impact: Maintain >95% similarity fidelity even after years of operation.

5. Knowledge Gap Detection (Proactive Curiosity)

Problem: The system is entirely reactive — it answers queries but never identifies what it doesn't know. True cognitive autonomy requires self-directed learning.

Solution: Temporal co-occurrence analysis — detect concepts that are frequently accessed close in time but have no synaptic connection, flagging them as knowledge gaps:

# Automatically runs hourly
gaps = engine.detect_knowledge_gaps(time_window_seconds=300)

# Returns structured insight:
# [
#   {
#     "concept_a": "Python asyncio event loop",
#     "concept_b": "FastAPI dependency injection",
#     "suggested_query": "How does asyncio relate to FastAPI dependency injection?",
#     "co_occurrence_count": 4
#   }
# ]

# Query endpoint
GET /insights/gaps?lookback_hours=24

# Fill gap manually (or via LLM agent)
POST /insights/fill-gap
{"concept_a_id": "mem_xxx", "concept_b_id": "mem_yyy",
 "explanation": "FastAPI uses asyncio's event loop internally..."}

The system becomes capable of saying what it doesn't understand and requesting clarification — the first step toward genuine cognitive autonomy.

Phase 4.4–4.5: Subconscious Daemon & LLM Integration

Subconscious Daemon (Autonomous Background Mind)

Phase 4.4 introduced SubconsciousAI — a worker that fires during idle cycles and calls an external LLM to generate synthetic dream memories: structured insights derived by reasoning over existing memory clusters, rather than through direct observation.

Phase 4.5 hardened this into a full SubconsciousDaemon — an independently managed asyncio process that orchestrates dream cycles, consolidation scheduling, and subconscious queue processing:

# The daemon is started automatically when the API starts up.
# It coordinates:
#   - Dream synthesis: SubconsciousAI → LLM → synthetic insights stored back
#   - Consolidation scheduling: ConsolidationWorker fired on a configurable interval
#   - Subconscious queue: novelty detection from Redis pub/sub stream

Configure in config.yaml:

haim:
  subconscious_ai:
    enabled: true
    api_url: "https://api.openai.com/v1/chat/completions"  # or Anthropic
    model: "gpt-4o-mini"
    # api_key: set via SUBCONSCIOUS_AI_API_KEY env var
    dream_interval_seconds: 300
    batch_size: 5

Dependency Injection Container

All major services (TierManager, AsyncRedisStorage, QdrantStore, SubconsciousAI, etc.) are now wired through src/mnemocore/core/container.py. This eliminates global singleton state and makes every subsystem testable in isolation:

from mnemocore.core.container import build_container

container = build_container(config)
engine   = container.engine()
tier_mgr = container.tier_manager()

LLM Agent Integration

src/mnemocore/llm_integration.py provides a high-level interface for attaching MnemoCore to any OpenAI/Anthropic-style LLM agent loop:

from mnemocore.llm_integration import MnemoCoreAgent

agent = MnemoCoreAgent(engine)

# Store agent observations
agent.observe("User prefers concise answers over verbose ones")

# Recall relevant context before a response
context = agent.recall("user preference", top_k=3)

API Reference (v2 compatibility)

This section documents the legacy HAIM REST service (mnemocore.api.main:app). It remains available for v2 clients but its global JSONL/tiering lifecycle is not AgentMemory-backed and must not be presented as v3 persistence. For v3, use create_v3_app(sqlite_path, scope_authorizer=...); the authorizer must approve the caller's complete requested scope before each operation.

Authentication

All endpoints require an API key via the X-API-Key header:

export HAIM_API_KEY="your-secure-key"
curl -H "X-API-Key: $HAIM_API_KEY" ...

Endpoints

`POST /store`

Store a new memory with optional context binding.

Request:
{
  "content": "FastAPI uses Pydantic v2 for request validation.",
  "metadata": {"source": "docs", "tags": ["python", "fastapi"]},
  "context": "ProjectAlpha",
  "agent_id": "agent-001",
  "ttl": 3600
}

Response:
{
  "ok": true,
  "memory_id": "mem_1739821234567",
  "message": "Stored memory: mem_1739821234567"
}

`POST /query`

Query memories by semantic similarity with optional context masking.

Request:
{
  "query": "How does FastAPI handle request validation?",
  "top_k": 5,
  "context": "ProjectAlpha"
}

Response:
{
  "ok": true,
  "query": "How does FastAPI handle request validation?",
  "results": [
    {
      "id": "mem_1739821234567",
      "content": "FastAPI uses Pydantic v2 for request validation.",
      "score": 0.8923,
      "metadata": {"source": "docs"},
      "tier": "hot"
    }
  ]
}

`POST /feedback`

Report outcome of a retrieved memory (Phase 4.0 reliability loop).

Request:
{
  "memory_id": "mem_1739821234567",
  "outcome": true,
  "comment": "This solution worked perfectly."
}

Response:
{
  "ok": true,
  "memory_id": "mem_1739821234567",
  "reliability_score": 0.714,
  "success_count": 4,
  "failure_count": 1
}

`GET /memory/{memory_id}`

Retrieve a specific memory with full metadata.

Response:
{
  "id": "mem_1739821234567",
  "content": "...",
  "metadata": {...},
  "created_at": "2026-02-17T20:00:00Z",
  "ltp_strength": 1.847,
  "epistemic_value": 0.73,
  "reliability_score": 0.714,
  "tier": "hot"
}

`DELETE /memory/{memory_id}`

Delete memory from all tiers and clean up synapses.

`POST /concept`

Define a symbolic concept for analogical reasoning.

{"name": "king", "attributes": {"gender": "man", "role": "ruler", "domain": "royalty"}}

`POST /analogy`

Solve analogies using VSA algebra: source:value :: target:?

Request:  {"source_concept": "king", "source_value": "man", "target_concept": "queen"}
Response: {"results": [{"value": "woman", "score": 0.934}]}

`GET /insights/gaps`

Detect knowledge gaps from recent temporal co-activity (Phase 4.0).

Response:
{
  "gaps_detected": 3,
  "knowledge_gaps": [
    {
      "concept_a": "asyncio event loop",
      "concept_b": "FastAPI middleware",
      "suggested_query": "How does event loop relate to middleware?",
      "co_occurrence_count": 5
    }
  ]
}

`POST /admin/consolidate`

Trigger manual semantic consolidation (normally runs automatically at 3 AM).

`GET /stats`

Engine statistics - tiers, synapse count, consolidation state. Requires X-API-Key by default.

`GET /health`

Public liveness and dependency diagnostics. A local-only runtime can report degraded when Redis is unavailable while remaining live.

`GET /ready`

Readiness of the initialized local runtime. Redis is reported diagnostically but is not required for the single-node prototype to become ready.

`GET /metrics/`

Prometheus metrics endpoint on the API listener (port 8100). The trailing slash is canonical.

Python Library Usage

For durable SQLite-backed memory with exact agent scopes, async/sync APIs, bitemporal recall, supersession, and evidence receipts, see the Persistent Agent Memory Quickstart. Retrieval in this foundation is SQLite FTS5 lexical search only; it is not semantic or embedding retrieval.

v3 scoped runtime

from mnemocore.agent_memory import AgentMemory, MemoryScope
from mnemocore.hybrid import HybridMemoryRuntime
import asyncio

async def main():
    scope = MemoryScope(tenant_id="local", user_id="demo", agent_id="assistant")
    async with await AgentMemory.open("memory.db", scope=scope) as memory:
        runtime = HybridMemoryRuntime(memory, scope=scope)
        await memory.remember("User prefers concise answers")
        results = await runtime.recall(scope, "concise", limit=3)
        print(results)

asyncio.run(main())

In v3, LiteEngine and the entire unscoped Memory(...) facade are removed and fail with a migration error; use AgentMemory with an explicit MemoryScope. Every runtime operation uses the complete scope tuple; there is no broader scope fallback. HAIMEngine remains available only for v2 compatibility: its global JSONL/tiering lifecycle is not AgentMemory-backed and must not be used as v3 persistence. Use HybridMemoryRuntime (or create_v3_app for HTTP) for v3. For a temporary bridge of its legacy store/query/delete_memory shape, use the deprecated HAIMEngineAdapter with an already-open, scope-bound AgentMemory client.

The hybrid runtime also has an explicit synchronous facade for synchronous applications. Do not call it from an already-running event loop; use the async runtime there. See the v3 hybrid-runtime guide for retrieval metadata, cognitive-plan safeguards, API composition, and the staged remaining work.

Advanced / Low-level (v2 compatibility only)

HAIMEngine below is retained for existing v2 integrations. It is deprecated for new work, does not enforce MemoryScope, and is not the v3 persistence layer. New v3 code must use the scoped runtime shown above.

from mnemocore.core.engine import HAIMEngine

engine = HAIMEngine(persist_path="./data/memory.jsonl")

# Store memories
engine.store("Python generators are lazy iterators", metadata={"topic": "python"})
engine.store("Use 'yield' to create generator functions", metadata={"topic": "python"})
engine.store("Redis XADD appends to a stream", goal_id="infrastructure")

# Query (global)
results = engine.query("How do Python generators work?", top_k=3)
for mem_id, score in results:
    mem = engine.get_memory(mem_id)
    print(f"[{score:.3f}] {mem.content}")

# Query with context masking
results = engine.query("data streams", top_k=5, context="infrastructure")

engine.close()

Analogical Reasoning

# Define concepts
engine.define_concept("king",  {"gender": "man",   "role": "ruler"})
engine.define_concept("queen", {"gender": "woman", "role": "ruler"})
engine.define_concept("man",   {"gender": "man"})

# VSA analogy: king:man :: ?:woman → queen
result = engine.reason_by_analogy(
    src="king", val="man", tgt="woman"
)
print(result)  # [("queen", 0.934), ...]

Working with the Binary HDV Layer Directly

from mnemocore.core.binary_hdv import BinaryHDV, TextEncoder, majority_bundle

encoder = TextEncoder(dimension=16384)

# Encode text
python_vec  = encoder.encode("Python programming")
fastapi_vec = encoder.encode("FastAPI framework")
error_vec   = encoder.encode("error handling")

# Bind concept to role
python_in_fastapi = python_vec.xor_bind(fastapi_vec)

# Bundle multiple concepts into prototype
web_dev_prototype = majority_bundle([python_vec, fastapi_vec, error_vec])

# Similarity
print(python_vec.similarity(web_dev_prototype))  # High (part of bundle)
print(python_vec.similarity(error_vec))          # ~0.5 (unrelated)

# Batch nearest-neighbor search
from mnemocore.core.binary_hdv import batch_hamming_distance
import numpy as np

database = np.stack([v.data for v in [python_vec, fastapi_vec, error_vec]])
distances = batch_hamming_distance(python_vec, database)

Reliability Feedback Loop

mem_id = engine.store("Always use asyncio.Lock() in async code, not threading.Lock()")
results = engine.query("async locking")

# It works — report success
engine.provide_feedback(mem_id, outcome=True, comment="Solved deadlock issue")

# Over time, high-reliability memories get 'verified' tag
# and are ranked above unproven ones in future queries

Semantic Consolidation

stats = engine.trigger_consolidation()
print(f"Created {stats['abstractions_created']} semantic anchors")
print(f"Consolidated {stats['memories_consolidated']} episodic memories")

# Automatic: runs every night at 3 AM via background asyncio task

Installation

Prerequisites

Python 3.10+
Redis and Qdrant are optional runtime services for legacy tiering and vector integrations. They are not required for local AgentMemory.
The current 2.x wheel retains its compatibility dependencies. It does not connect to Redis or Qdrant merely by importing mnemocore.agent_memory.

Quick Start (AgentMemory — stable local-first path)

# 1. Install. No Redis/Qdrant service is needed for AgentMemory.
pip install -e .

# 2. Use AgentMemory with an explicit scope and local SQLite store.
from pathlib import Path
from mnemocore.agent_memory import MemoryScope, SyncAgentMemory

scope = MemoryScope(tenant_id="local", user_id="robin", agent_id="assistant")
memory = SyncAgentMemory.open(Path("agent-memory.sqlite3"), scope=scope)
memory.remember("User prefers concise answers")
print(memory.recall("concise"))

See "Full server setup (optional)" below for HAIM, Redis, Docker, and the REST API.

Full server setup (optional, for Redis-backed etc.)

To include dev tools (pytest, mypy, black, etc.):

pip install -e ".[dev]"

Start Redis

docker run -d -p 6379:6379 redis:7.2-alpine

Set API key (never hardcode — use env var or .env file)

Windows PowerShell:

$env:HAIM_API_KEY = "your-secure-key-here"

Linux / macOS:

export HAIM_API_KEY="your-secure-key-here"

Start the API locally

PowerShell; use export MNEMOCORE_HOST=127.0.0.1 on Unix

$env:MNEMOCORE_HOST = "127.0.0.1" mnemocore

The API is now live at http://localhost:8100. Visit http://localhost:8100/docs for the interactive Swagger UI.

Do not bind MnemoCore directly to a public IP unless it is behind a trusted reverse proxy, TLS, and a strong HAIM_API_KEY. Diagnostics (/stats, /rate-limits, /metrics) require X-API-Key by default.

Using the .env file

Copy the provided template and fill in your values — the API and docker-compose both pick it up automatically:

cp .env.example .env
# Edit .env and set HAIM_API_KEY, REDIS_PASSWORD, QDRANT_API_KEY, etc.

Note: .env is listed in .gitignore and must never be committed. Only .env.example (with placeholder values) belongs in version control.

Full Stack with Docker Compose

# Requires .env with HAIM_API_KEY, REDIS_PASSWORD, and QDRANT_API_KEY set.
docker compose config
docker compose up -d

This starts MnemoCore, Redis 7.2, and Qdrant. Only the API is published, on 127.0.0.1:8100; metrics are served from /metrics/ on the same port.

With Qdrant (Phase 4.x Scale)

# Start Qdrant alongside Redis
docker run -d -p 6333:6333 qdrant/qdrant

# Enable in config.yaml
qdrant:
  enabled: true
  host: localhost
  port: 6333

Configuration

All configuration lives in config.yaml. Sensitive values can be overridden with environment variables — the config loader looks for HAIM_-prefixed vars and also honours per-service overrides like HAIM_API_KEY, REDIS_PASSWORD, QDRANT_API_KEY, HAIM_CORS_ORIGINS, and SUBCONSCIOUS_AI_API_KEY.

haim:
  version: "5.0"
  dimensionality: 16384        # Binary vector dimensions (must be multiple of 64)

  encoding:
    mode: "binary"             # "binary" (recommended) or "float" (legacy, deprecated)
    token_method: "bundle"     # "bundle" (XOR+permute) or "hash"

  tiers:
    hot:
      max_memories: 2000       # Max nodes in RAM
      ltp_threshold_min: 0.7   # Evict below this LTP strength
      eviction_policy: "lru"
    warm:
      max_memories: 100000     # Max nodes in Redis/mmap
      ltp_threshold_min: 0.3
    cold:
      storage_backend: "filesystem"   # "filesystem" or "s3"
      compression: "gzip"

  ltp:
    initial_importance: 0.5
    decay_lambda: 0.01         # Higher = faster forgetting
    permanence_threshold: 0.95 # LTP above this is immune to decay
    half_life_days: 30.0

  hysteresis:
    promote_delta: 0.15        # LTP must exceed threshold by this much to promote
    demote_delta: 0.10

  redis:
    url: "redis://localhost:6379/0"
    stream_key: "haim:subconscious"
    max_connections: 10
    socket_timeout: 5
    # password: set via REDIS_PASSWORD env var

  qdrant:
    url: "http://localhost:6333"
    collection_hot:  "haim_hot"
    collection_warm: "haim_warm"
    enabled: false
    # api_key: set via QDRANT_API_KEY env var

  security:
    # api_key: set via HAIM_API_KEY env var — never hardcode here
    cors_origins: ["http://localhost:3000"]
    public_health_enabled: true
    public_stats_enabled: false
    public_rate_limits_enabled: false
    public_metrics_enabled: false
    require_api_key_for_export: true

  subconscious_ai:
    enabled: false
    api_url: "https://api.openai.com/v1/chat/completions"
    model: "gpt-4o-mini"
    dream_interval_seconds: 300
    batch_size: 5
    # api_key: set via SUBCONSCIOUS_AI_API_KEY env var

  observability:
    metrics_port: 8100
    log_level: "INFO"
    structured_logging: true

  paths:
    data_dir: "./data"
    memory_file: "./data/memory.jsonl"
    codebook_file: "./data/codebook.json"
    concepts_file: "./data/concepts.json"
    synapses_file: "./data/synapses.json"
    warm_mmap_dir: "./data/warm_tier"
    cold_archive_dir: "./data/cold_archive"

  mcp:
    enabled: false
    transport: "stdio"
    host: "127.0.0.1"
    port: 8110
    api_base_url: "http://localhost:8100"

Security Note

MnemoCore requires an explicit API key. There is no default fallback key in production builds.

# Generate a cryptographically secure key:
python -c "import secrets; print(secrets.token_urlsafe(32))"

# Set it (never commit this value):
export HAIM_API_KEY="<generated-value>"

Persistent webhooks store only an opaque secret_ref; the signing secret is resolved when a delivery is sent. Inline signing secrets, legacy plaintext records, and all custom headers are rejected for persistent webhook managers. Writes are serialized and atomically replaced, and malformed persistence loads fail closed without partial in-memory state. Memory-only webhook managers retain the legacy inline-secret/header behavior for compatibility.

AgentMemory also exposes a physical SQLite erase operation with exact-scope ownership checks, optional whole-supersession-component cascade, dependent-row cleanup, database integrity checks, and a content-free ErasureReceipt. All database users must cooperate with the store's sidecar lock; raw SQLite clients are outside the safety contract during erasure. Backups, derived external artifacts, power-loss recovery, and exhaustive failure injection remain separate production gates.

MCP Server Integration (v2 compatibility)

MnemoCore exposes a legacy Model Context Protocol (MCP) server for existing v2 integrations. It proxies the legacy HAIM REST API and is not backed by v3 AgentMemory truth. New scoped integrations should use the AgentMemory MCP helpers in mnemocore.integrations.mcp (or build on HybridMemoryRuntime).

Setup

# Start the legacy v2 API first
uvicorn mnemocore.api.main:app --host 0.0.0.0 --port 8100

# Configure MCP in config.yaml
haim:
  mcp:
    enabled: true
    transport: "stdio"  # or "sse" for streaming

# Run MCP server
python -m mnemocore.mcp.server

Claude Desktop Configuration

Add to your Claude Desktop config.json:

{
  "mcpServers": {
    "mnemocore": {
      "command": "python",
      "args": ["-m", "mnemocore.mcp.server"],
      "env": {
        "HAIM_API_KEY": "your-key",
        "HAIM_BASE_URL": "http://localhost:8100"
      }
    }
  }
}

Once connected, the agent can:

store_memory(content, context) — persist learned information
query_memory(query, context, top_k) — recall relevant memories
provide_feedback(memory_id, outcome) — signal what worked
get_knowledge_gaps() — surface what it doesn't understand

Observability

MnemoCore ships with built-in Prometheus metrics and structured logging.

Prometheus Metrics

Available at canonical GET /metrics/ on the API listener, port 8100:

Metric	Description
`haim_api_request_count`	Total requests by endpoint and status
`haim_api_request_latency_seconds`	Request latency histogram
`haim_storage_operation_count`	Store/query/delete operations
`haim_hot_tier_size`	Current HOT tier memory count
`haim_synapse_count`	Active synaptic connections

Grafana Dashboard

A sample Grafana dashboard config is available at grafana-dashboard.json in the repository root. Import it directly into Grafana via Dashboards → Import → Upload JSON file.

Structured Logging

All components use structured Python logging with contextual fields:

2026-02-17 20:00:00 INFO  Stored memory mem_1739821234567 (EIG: 0.7823)
2026-02-17 20:00:01 INFO  Memory mem_1739821234567 reliability updated: 0.714 (4✓ / 1✗)
2026-02-17 03:00:00 INFO  Consolidation complete: abstractions_created=12, consolidated=847
2026-02-17 04:00:00 INFO  Knowledge gap detected: asyncio ↔ FastAPI middleware (5 co-occurrences)

Testing

# Run full test suite
pytest

# Run with coverage
pytest --cov=src --cov-report=html

# Run specific feature tests
pytest tests/test_xor_attention.py         # Contextual masking
pytest tests/test_stability.py             # Reliability/Bayesian stability
pytest tests/test_consolidation.py         # Semantic consolidation
pytest tests/test_engine_cleanup.py        # Cleanup and decay
pytest tests/test_phase43_regressions.py   # Phase 4.3 regression guardrails
pytest tests/test_tier_manager.py          # Tier demotion / promotion logic
pytest tests/test_dream_loop.py            # Subconscious dream loop
pytest tests/test_subconscious_ai_worker.py # LLM-powered dream worker (if offline: uses mocks)
pytest tests/test_recursive_synthesizer.py  # Deep concept synthesis
pytest tests/test_batch_ops.py             # Bulk ingestion operations
pytest tests/test_mcp_server.py            # MCP server adapter

# End-to-end flow
pytest tests/test_e2e_flow.py -v

Roadmap

v3.0.0 beta: released surface

Exact-scope AgentMemory persistence, history, timeline, and context compilation
Hybrid lexical/BinaryHDV retrieval with versioned deterministic scoring
Rebuildable tier and graph projections
Validated, atomic cognitive plan application with provenance
Scoped HTTP composition that fails closed without authorization
Explicit migration away from LiteEngine and implicit Memory(...)

v2 compatibility inventory

Binary HDV core (XOR bind / bundle / permute / Hamming)
Three-tier HOT/WARM/COLD memory lifecycle
Async API + MCP integration
XOR attention masking + Bayesian reliability updates
Semantic consolidation, immunology cleanup, and gap detection/filling
Temporal recall: episodic chaining + chrono-weighted query
Phase 4.4 — Subconscious AI Worker (LLM-powered dream synthesis)
Phase 4.5 — Subconscious Daemon, persistence hardening, tier-manager demotion race fix
Dependency-injection Container pattern (replaces singleton)
HNSW in-process index for hot-tier ANN search
Batch operations for bulk ingestion
Meta-cognition layer: GoalTree + LearningJournal
Phase 5 — Cognitive Architecture: Working Memory, Episodic/Semantic/Procedural stores, Meta Memory, Self-Improvement Worker, 7-phase Pulse loop
Phase 5 Hardening — 136 additional tests, config organization
Phase 6 — Research Services: StrategyBank, KnowledgeGraph, MemoryScheduler, SAMEP, 4 new Pulse phases
1291+ tests passing

Next Steps

Production hardening: distributed HOT-tier, CUDA kernels, chaos engineering
Self-improvement Phase 1: enable writes with human-in-the-loop approval
Multi-modal memory: image/audio embedding via CLIP/Whisper
Additional first-party integration adapters
Extended observability standardization (mnemocore_* metric prefix)

Contributing

MnemoCore is an active research project. Contributions are welcome — especially:

Performance: CUDA kernels, FAISS integration, async refactoring
Algorithms: Better clustering for consolidation, improved EIG formulas
Integrations: New storage backends, LLM connectors
Tests: Coverage for edge cases, property-based testing

Process

# Fork and clone
git checkout -b feature/your-feature-name

# Make changes, ensure tests pass
pytest

# Commit with semantic message
git commit -m "feat(consolidation): add LLM-powered prototype labeling"

# Open PR — describe the what, why, and performance impact

Please follow the implementation patterns established in docs/ARCHITECTURE.md and docs/ROADMAP.md for architectural guidance, and review CHANGELOG.md to understand what has already landed.

Documentation

Document	Description
docs/ARCHITECTURE.md	System architecture with Mermaid diagrams
docs/API.md	Complete REST API reference (41 endpoints)
docs/CLI.md	Command-line interface reference
docs/CONFIGURATION.md	All configuration options (41 dataclasses)
docs/DEPLOYMENT.md	Docker, Compose, and Kubernetes deployment
docs/EVENTS.md	Event system and webhook delivery
docs/FORGETTING.md	SM-2 spaced repetition and decay system
docs/LLM_INTEGRATION.md	Multi-provider LLM integration
docs/GLOSSARY.md	Terms and abbreviations
docs/SUBCONSCIOUS_AI.md	Subconscious worker and dream pipeline
docs/RECONSTRUCTIVE_MEMORY.md	Reconstructive recall module
docs/EMBEDDING_REGISTRY.md	Embedding model versioning
CHANGELOG.md	Release history
CONTRIBUTING.md	Contribution guidelines

License

MIT License — see LICENSE for details.

Contact

Robin Granberg
📧 robin@veristatesystems.com

Building the cognitive substrate for the next generation of autonomous AI.

Downloads last month: 10

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support