Cell-mem

Brain-inspired memory system for AI Agents — an MCP (Model Context Protocol) server that gives AI agents persistent, multi-layered memory with consolidation, self-reflection, generative replay, and creative hypothesis discovery.

Cell-mem models the human brain's memory architecture: four interconnected memory layers operating at different timescales, governed by neuro-inspired consolidation and forgetting processes. It ships as an MCP server — drop it into Claude Code, Codex CLI, or any MCP-compatible agent host.

Status: Stable — see CHANGELOG.md for version history.

---

Architecture

MCP Server (stdio + HTTP transport)
│
├── 12 MCP Tools
│   ├── memory_save            — Store a memory
│   ├── memory_recall          — Cross-layer retrieval
│   ├── memory_status          — System health dashboard
│   ├── memory_associate       — Link two memories (graph edge)
│   ├── memory_forget          — Manual memory removal
│   ├── memory_consolidate     — Trigger consolidation cycle
│   ├── memory_verify          — Check falsifiable conditions
│   ├── memory_reflect         — Self-reflection (failure analysis + strategy eval)
│   ├── memory_replay          — Trigger generative replay (hypothesis creation)
│   ├── memory_hypothesis_feedback — Confirm/reject a creative hypothesis
│   ├── memory_creative_pool   — Inspect the hypothesis pool
│   └── memory_check_environment  — Detect environment changes → auto-verify
│
├── Memory Layers (brain-inspired)
│   ├── Working Memory    <minutes>  ~50 items, attention-based decay
│   ├── Episodic Memory   <days>     pattern-separated experience storage
│   ├── Semantic Memory   <months>   facts with falsifiable conditions
│   └── Procedural Memory <months>   skill/strategy templates with RL weighting
│
├── Consolidation Processor
│   ├── Emotional scoring (multi-dimensional: recency, frequency, valence, surprise)
│   ├── DBSCAN pattern detection
│   ├── Forgetting (low-score → cold storage archive, rescuer support)
│   └── State persistence across restarts
│
├── Reflective System
│   ├── Effect attribution — "What went wrong and why?"
│   ├── Strategy evaluation — Success trends, better variants, redundancy
│   ├── Knowledge gap detection — Missing info? Retrieval failure?
│   └── Result processing — Update procedural weights, adjust semantic confidence
│
├── Generative Replay Engine
│   ├── 5-stage algorithm: biased sampling → random walk → cross-domain pairing
│   │                      → 4-layer noise filter → creative pool management
│   ├── Creative pool: hypothesis lifecycle (pending → confirmed/rejected → promoted)
│   └── 10 noise constraints to prevent hallucinations from persisting
│
└── Storage (SQLite)
    ├── sqlite-vec vector search (384d all-MiniLM-L6-v2 embeddings)
    ├── FTS5 full-text search with OR semantics
    ├── Graph store (NetworkX-backed, spreading activation)
    └── Cold storage archive (forgotten but rescuable)

---

Quick Start

Installation

# From the repository root
pip install -e .

# With HTTP transport support
pip install -e ".[http]"

# With development tools (linting, testing)
pip install -e ".[dev]"

Run as MCP Server

# stdio mode — agent launches as subprocess (no network)
python -m cell_mem.server

# HTTP mode — daemon for hook scripts, multiple agents
python -m cell_mem.server --http --port 8765

# HTTP with shared-secret authentication (recommended for production)
python -m cell_mem.server --http --port 8765 --api-key "your-secret-here"

# Preload embedding model (avoids ~30s first-request delay)
python -m cell_mem.server --preload

# With seed knowledge (pre-populate semantic memory)
python -m cell_mem.server --seed-config config/seed_knowledge.example.json

MCP Client Configuration

Add to your agent's MCP configuration:

Claude Code (~/.claude/claude_desktop_config.json):

{
  "mcpServers": {
    "cell-mem": {
      "command": "python",
      "args": ["-m", "cell_mem.server", "--db", "/path/to/cell_mem.db"]
    }
  }
}

Codex CLI:

{
  "mcpServers": {
    "cell-mem": {
      "command": "python",
      "args": ["-m", "cell_mem.server"],
      "env": { "CELL_MEM_DB": "/path/to/cell_mem.db" }
    }
  }
}

---

Memory Layers

Working Memory (seconds–minutes)

• Capacity-limited (~50 items), attention-based decay
• Items are pushed to a "preheated zone" before aging out
• Emulates the prefrontal cortex's short-term buffer

Episodic Memory (hours–days)

• Pattern-separated storage: 384d content embedding → 2048d projection
• Reduces interference between similar but distinct episodes
• Consolidation scoring determines retention priority

Semantic Memory (weeks–months)

• Facts, knowledge, and rules with optional falsifiable conditions
• Conditions define what would make the fact outdated (e.g., "package.json version changed")
• memory_verify checks conditions against environment snapshots
• High-confidence + locked lifecycle → resist unlearning

Procedural Memory (days–months)

• Skill/strategy templates triggered by cosine similarity to current context
• Reinforcement learning: success → weight × 1.05, failure → weight × 0.85
• Explore/exploit balance: 80% exploit (best match), 20% explore (novel picks)
• Templates with weight < 0.25 → candidates for reflection review

---

Key Mechanisms

Consolidation

Automatic (via should_run()) or manual (memory_consolidate) cycles:

1. Score all episodes on recency, frequency, emotional valence, surprise
2. Identify low-score candidates for forgetting
3. After 3 consecutive low-score cycles → archive to cold storage (rescuable)
4. Run DBSCAN pattern clustering to detect emerging knowledge patterns

Self-Reflection

Four-dimensional meta-reasoning over failure events:

• Dimension 1 — Effect Attribution: Causal analysis of failures
• Dimension 2 — Strategy Evaluation: Success rate trends, variant comparison
• Dimension 3 — Knowledge Gap Detection: Missing facts or retrieval failures
• Dimension 4 — Result Processing: Update procedural weights, adjust confidences, create meta-knowledge

Generative Replay

Five-stage creative hypothesis engine inspired by hippocampal replay:

1. Biased sampling — pick K=3 seeds proportional to recency × emotional salience × novelty
2. Random walk — L=3 steps per seed, 80/20 strong/weak edge sampling
3. Cross-domain pairing — pair low-similarity concepts from different seeds
4. 4-layer noise filter — contradiction check, triviality filter, dual-source verification, stability requirement
5. Creative pool management — 10 noise constraints, pending → confirmed → promoted lifecycle

Falsifiable Conditions

Each semantic fact can carry a falsifiable_condition:

{
  "field": "package.json",
  "operator": "value_changed",
  "value": "react"
}

memory_check_environment compares current vs. last snapshot → auto-triggers memory_verify for affected facts. Or use memory_verify manually with a specific fact ID.

---

Python API

from cell_mem import MemorySystem

# Initialize (all layers + embedding model)
ms = MemorySystem("cell_mem.db")

# Store across layers
ms.save("User prefers dark theme", memory_type="semantic", confidence=0.9)
ms.save("Fixed the login bug with OAuth", memory_type="episodic")
ms.save("When encountering CORS errors, check server middleware first",
        memory_type="procedural", trigger_condition="CORS error debugging")

# Recall (cross-layer: semantic FTS5 + episodic embedding + procedural context)
results = ms.recall("How to debug CORS?")

# Graph associations
ms.associate(id_a, id_b, weight=0.8, relation="related_to")

# Status dashboard
status = ms.status()
# Layers: working/episodic/semantic/procedural counts + consolidation stats
# + creative pool + LLM usage + reflection history

# Self-reflection
ms.reflect(task="Fix CORS bug", outcome="Failure", dimensions="all")

# Generative replay (auto-creates hypotheses from memory graph)
ms.replay(theme_text="frontend debugging")

# Hypothesis feedback (confirmed → confidence boost; rejected → ignore_count++)
ms.record_hypothesis_feedback("hyp_abc123", confirmed=True)

# Environment change detection → auto-verify
ms.check_environment({"node_version": "18", "react_version": "19.0"})

ms.shutdown()

LLM Backend Configuration

LLM-powered features (reflection, replay, emotional scoring) can optionally use an LLM:

ms = MemorySystem(
    "cell_mem.db",
    llm_backend="openai",       # "openai" or "claude"
    llm_api_key="sk-...",       # or set OPENAI_API_KEY / ANTHROPIC_API_KEY env var
    llm_daily_limit=100,        # rate limiting (default: 100 calls/day)
)

Without an LLM, emotional scoring falls back to rule-based heuristics, and reflection/replay operations return informative errors. Core save/recall/status do not require an LLM.

---

Project Structure

src/cell_mem/
├── __init__.py              # Public API exports
├── models.py                # Pydantic data models
├── memory_system.py         # Top-level facade (main API)
├── server.py                # MCP server entry point
│
├── storage/                 # SQLite + vector storage
│   ├── sqlite_store.py      # Schema, migrations, meta table
│   ├── vector_store.py      # sqlite-vec and ChromaDB backends
│   └── search.py            # FTS5 search engine
│
├── embedding/               # Embedding models
│   └── local.py             # SentenceTransformers (all-MiniLM-L6-v2)
│
├── memory/                  # Four memory layers
│   ├── working.py           # Working memory (attention decay)
│   ├── episodic.py          # Episodic memory (pattern separation)
│   ├── semantic.py          # Semantic memory (falsifiable facts)
│   └── procedural.py        # Procedural memory (RL-weighted templates)
│
├── graph/                   # Associative graph
│   ├── store.py             # NetworkX graph store
│   ├── activation.py        # Spreading activation retrieval
│   └── networkx_store.py    # NetworkX adapter
│
├── consolidation/           # Sleep-like consolidation
│   ├── scorer.py            # Multi-dimension episode scoring
│   ├── detector.py          # DBSCAN pattern detection
│   ├── emotional.py         # Emotional valence evaluation
│   └── scheduler.py         # Cycle orchestrator + forgetting
│
├── reflection/              # Meta-reasoning
│   └── engine.py            # 4-dimension reflection engine
│
├── conditions/              # Falsifiable conditions
│   └── evaluator.py         # Condition checking + environment snapshots
│
├── replay/                  # Generative replay
│   ├── engine.py            # 5-stage replay algorithm
│   └── creative_pool.py     # Hypothesis lifecycle management
│
├── llm/                     # LLM abstraction
│   ├── client.py            # Base client + rate limiter
│   └── backends.py          # OpenAI + Claude backends (stdlib only)
│
└── tools/                   # MCP tool registrations
    ├── save.py              # memory_save
    ├── recall.py            # memory_recall
    ├── status.py            # memory_status
    ├── verify.py            # memory_verify
    ├── reflect.py           # memory_reflect
    ├── replay.py            # memory_replay + creative pool tools
    └── stubs.py             # memory_associate, forget, consolidate, tool wiring

---

Design Principles

• Zero new pip dependencies for core operations. LLM calls use stdlib urllib only. Dependencies (mcp, sentence-transformers, numpy, networkx, scikit-learn) are all well-established packages.
• SSRF protection. All LLM API calls validate the target URL against blocked private IP ranges (RFC 1918, link-local, CGNAT, IPv6 private).
• SQL-first architecture. SQLite with WAL mode, FTS5, sqlite-vec — all data local, no external services required.
• Brain-inspired, not brain-simulated. Algorithms are inspired by neuroscience (pattern separation, spreading activation, hippocampal replay) but optimized for practical agent memory, not biological fidelity.
• Graceful degradation. Optional features (LLM, HTTP, ChromaDB) degrade cleanly when not configured. Core memory operations always work.

---

Known Limitations

• Embedding model first load. First startup downloads all-MiniLM-L6-v2 (~90 MB) and takes ~30 seconds. Use --preload flag to warm up at startup.
• sqlite-vec requires Rust toolchain for compilation from source. On most platforms, pre-built wheels are available via pip. If building from source, install Rust from rustup.rs.
• API key via CLI is visible in process lists on multi-user systems. Prefer environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY) for production deployments.
• No MCP tool rate limiting. LLM calls are rate-limited (default 100/day), but MCP tools themselves have no per-call throttle. In local agent deployments this is not a practical concern.
• Preference pipeline needs LLM for optimal extraction. Keyword-based fallback works without LLM, but extraction quality improves significantly with an LLM configured.

---

Requirements

• Python ≥ 3.11
• SQLite ≥ 3.35 (for sqlite-vec support)
• Optional: OpenAI or Anthropic API key (for LLM-powered features)

---

License

MIT — see LICENSE for full text.

Copyright (c) 2026 Siqi Liu