shodh-memory 0.1.5

We built this because AI agents forget everything between sessions. They make the same mistakes, ask the same questions, lose context constantly.

Shodh-Memory fixes that. It's a cognitive memory system—Hebbian learning, activation decay, semantic consolidation—packed into a single 8MB binary that runs offline.

How it works:

Experiences flow through three tiers based on Cowan's working memory model [1]. New information enters capacity-limited working memory, overflows into session storage, and consolidates into long-term memory based on importance. When memories are retrieved together successfully, their connections strengthen—classic Hebbian learning [2]. After enough co-activations, those connections become permanent. Unused memories naturally fade.

Working Memory ──overflow──▶ Session Memory ──importance──▶ Long-Term Memory
   (100 items)                  (500 MB)                      (RocksDB)

Architecture

Storage & Retrieval

• Vamana graph index for approximate nearest neighbor search [3]
• MiniLM-L6 embeddings (384-dim, 25MB) for semantic similarity
• TinyBERT NER (15MB) for entity extraction (Person, Organization, Location, Misc)
• RocksDB for durable persistence across restarts

Cognitive Processing

• Named entity recognition — TinyBERT extracts entities; boosts importance and enables graph relationships
• Spreading activation retrieval — queries activate related memories through semantic and graph connections [5]
• Activation decay — exponential decay A(t) = A₀ · e^(-λt) applied each maintenance cycle
• Hebbian strengthening — co-retrieved memories form graph edges; weight increases on co-activation
• Long-term potentiation — edges surviving threshold co-activations become permanent

Semantic Consolidation

• Episodic memories older than 7 days compress into semantic facts
• Entity extraction preserves key information during compression

Use cases

Local LLM memory — Give Claude, GPT, or any local model persistent memory across sessions.

Robotics & drones — On-device experience accumulation without cloud round-trips.

Edge AI — Run on Jetson, Raspberry Pi, industrial PCs. Sub-millisecond retrieval, zero network dependency.

Personal knowledge base — Your own searchable memory. Decisions, learnings, discoveries—private and local.

Compared to alternatives

Performance

Measured on Intel i7-1355U (10 cores, 1.7GHz), release build.

API Latencies

Knowledge Graph (Criterion benchmarks)

Neural Models

Installation

Claude Code / Claude Desktop:

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "shodh-memory": {
      "command": "npx",
      "args": ["-y", "@shodh/memory-mcp"]
    }
  }
}

Config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Python:

pip install shodh-memory

From source:

cargo build --release
./target/release/shodh-memory-server

Usage

Python

from shodh_memory import Memory

memory = Memory(storage_path="./my_data")

# Store
memory.remember("User prefers dark mode", memory_type="Decision")
memory.remember("JWT tokens expire after 24h", memory_type="Learning")

# Search
results = memory.recall("user preferences", limit=5)

# Get memory statistics
stats = memory.get_stats()

REST API

# Store
curl -X POST http://localhost:3030/api/remember \
  -H "Content-Type: application/json" \
  -H "X-API-Key: your-key" \
  -d '{"user_id": "agent-1", "content": "Deployment requires Docker 24+", "tags": ["deployment"]}'

# Search
curl -X POST http://localhost:3030/api/recall \
  -H "Content-Type: application/json" \
  -H "X-API-Key: your-key" \
  -d '{"user_id": "agent-1", "query": "deployment requirements", "limit": 5}'

Memory types

Different types get different importance weights in the scoring model:

• Decision (+0.30) — choices, preferences, conclusions
• Learning (+0.25) — new knowledge, facts learned
• Error (+0.25) — mistakes, things to avoid
• Discovery, Pattern (+0.20) — findings, recurring behaviors
• Task (+0.15) — work items
• Context, Observation (+0.10) — general info

API reference

Python client (API parity with REST)

REST endpoints

All protected endpoints require X-API-Key header.

Authentication

# Development mode (SHODH_API_KEYS not set)
curl -H "X-API-Key: sk-shodh-dev-4f8b2c1d9e3a7f5b6d2c8e4a1b9f7d3c" ...

# Production mode (required)
export SHODH_API_KEYS="your-secure-key-1,your-secure-key-2"
export SHODH_ENV=production

Configuration

SHODH_PORT=3030                    # Default: 3030
SHODH_MEMORY_PATH=./data           # Default: ./shodh_memory_data
SHODH_API_KEYS=key1,key2           # Required in production
SHODH_MAINTENANCE_INTERVAL=300     # Decay cycle (seconds)
SHODH_ACTIVATION_DECAY=0.95        # Decay factor per cycle

Platform support

References

[1] Cowan, N. (2010). The Magical Mystery Four: How is Working Memory Capacity Limited, and Why? Current Directions in Psychological Science, 19(1), 51-57. https://pmc.ncbi.nlm.nih.gov/articles/PMC4207727/

[2] Magee, J.C., & Grienberger, C. (2020). Synaptic Plasticity Forms and Functions. Annual Review of Neuroscience, 43, 95-117. https://pmc.ncbi.nlm.nih.gov/articles/PMC10410470/

[3] Subramanya, S.J., et al. (2019). DiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node. NeurIPS 2019. https://papers.nips.cc/paper/9527-diskann-fast-accurate-billion-point-nearest-neighbor-search-on-a-single-node

[4] Dudai, Y., Karni, A., & Born, J. (2015). The Consolidation and Transformation of Memory. Neuron, 88(1), 20-32. https://pmc.ncbi.nlm.nih.gov/articles/PMC4183265/

[5] Anderson, J.R. (1983). A Spreading Activation Theory of Memory. Journal of Verbal Learning and Verbal Behavior, 22(3), 261-295.

License

Apache 2.0

MCP Registry · PyPI · npm · GitHub