Synaptic (fresh) — AI Memory Store

Synaptic is a local, cacheable memory system designed to reduce context/token spend by turning valuable thoughts, decisions, constraints, and insights into small “memory atoms” that can be retrieved, strengthened, decayed, and pruned like synapses.

This repo is intentionally not trading-specific. It's a general-purpose AI memory layer.

Core idea

• Atoms (neurons): small records worth keeping.
• Strength (synapse): the more an atom is used (retrieved/referenced), the more it strengthens.
• Decay: unused atoms slowly lose priority.
• Pruning: when space is constrained, weak / stale atoms are evicted.
• L1 Retrieval: find the most relevant atoms for a query.
• L2 Discovery: expand through neighbors + co-activation to surface adjacent ideas and propose new patterns (meta-atoms / clusters), similar in spirit to embeddings but richer.

Everything is artifact-first:

• append-only JSONL ledgers
• an SQLite index (FTS + edges)
• deterministic IDs + hashes

Data layout

By default, Synaptic stores data in ./synaptic_data/ (override with SYNAPTIC_HOME).

synaptic_data/
  atoms.jsonl                 # memory atoms (append-only)
  activations.jsonl           # retrieval/use events (append-only)
  synaptic.sqlite             # SQLite index: atoms + FTS + edges + coactivations
  blobs/                      # optional larger payloads (future)

Quick start

# from repo root
python -m synaptic.cli init

python -m synaptic.cli add --type idea --scope colony --tags orion,blob \
  --content "Guardian + TrustMeter gates are the civilizational control plane."

python -m synaptic.cli search "trust meter gates" --k 8
python -m synaptic.cli brief "what are our safety principles?" --k 10 --l2 8
python -m synaptic.cli prune --max-mb 50

Slash-command / local tool integration

Synaptic is designed to sit behind a local executor. If you want a strict JSON protocol, see:

• docs/COMMAND_PROTOCOL.md

Philosophy

• Truth > coherence: store receipts and provenance.
• Strong defaults: read-only behaviors; human approvals for destructive operations in external tools.
• Compression: meta-atoms emerge from repeated co-activation to reduce future context.

License

MIT (see LICENSE).

Decay + stable hashing

• Strength decays exponentially with a configurable half-life (SYNAPTIC_DECAY_HALF_LIFE_DAYS).
• Retrieval applies dynamic decay for ranking, and you can persist decay with syn decay (or --decay).
• The local hasher-embedder uses sha256-based stable hashing (deterministic across runs).

---

GitHub-ready notes

• Local memory storage is not meant to be committed. By default Synaptic writes to ./synaptic_data/.
  • Use SYNAPTIC_HOME to point storage elsewhere.
• CI is included via GitHub Actions: .github/workflows/ci.yml
• Console script: syn (after pip install -e .)

Quick install (editable)

python -m pip install -e .
syn init
syn add --type idea --scope colony --tags synaptic --content "Hello memory."
syn brief "what did I just add?" --k 6 --l2 4 --meta 2

Repository: https://github.com/resonantlabsai/synaptic.git