Hicortex

Self-improving long-term memory for AI agents. Capture sessions, distill lessons overnight, inject them on the next run. Works with Claude Code, OpenClaw, and any MCP-compatible agent.

Named after the hippocampus (fast encoding) and neocortex (slow consolidation) — the two brain systems that turn fleeting experiences into lasting knowledge.

Install

npx @gamaze/hicortex init

That's it. Auto-detects your environment, picks an LLM (Ollama / Claude CLI / API key), installs a local daemon (launchd on macOS, systemd on Linux), and registers MCP tools with Claude Code.

For multi-machine setups, point clients at a shared server:

npx @gamaze/hicortex init --server https://your-server.example.com

Full docs: hicortex.gamaze.com/docs

What it does

INGEST (nightly)             CONSOLIDATE (nightly)        RETRIEVE (instant)
┌──────────────────┐        ┌──────────────────────┐      ┌─────────────────────┐
│ Session transcripts        │ 1. Score importance  │      │ BM25 + vector search│
│ → LLM distillation         │    (local LLM)       │      │ → RRF fusion        │
│ → Local embedding          │ 2. Reflect & learn   │      │ → Graph traversal   │
│ → Store                    │    (cloud LLM)       │      │ → Composite scoring │
└──────────────────┘        │ 3. Auto-link by      │      │ → Strengthen on     │
                             │    vector similarity │      │    access           │
                             │ 4. Decay + prune     │      └─────────────────────┘
                             └──────────────────────┘
                                       ↓
                             Lessons (memory_type="lesson")
                                       ↓
                             Injected into CLAUDE.md / agent context

Memories decay slower the more important and frequently used they are, strengthen on retrieval, and are linked automatically to related memories. Retrieval is zero-LLM: BM25 full-text + vector search fused with Reciprocal Rank Fusion, scored by similarity (40%) + strength (30%) + connections (20%) + recency (10%).

MCP tools

Six MCP tools your agent can call:

Tool	Purpose
`hicortex_search`	Semantic search across all stored memories
`hicortex_context`	Recent decisions + project state for the current session
`hicortex_ingest`	Store a memory directly
`hicortex_lessons`	Actionable lessons from nightly reflection
`hicortex_update`	Fix incorrect memories (re-embeds on content change)
`hicortex_delete`	Remove memories with cascade cleanup

Plus skills: /learn to save explicit learnings.

Stack

TypeScript, Node.js 18+
better-sqlite3 + sqlite-vec + FTS5 (semantic + full-text search in one DB)
@huggingface/transformers (bge-small-en-v1.5 ONNX, runs on CPU)
MCP protocol over HTTP/SSE (Claude Code, OpenClaw, any MCP client)
Multi-provider LLM — Ollama, Claude CLI, OpenAI, Anthropic, Google, OpenRouter, or any OpenAI-compatible endpoint
Auto-detects Ollama models, Claude CLI, API keys during setup

Architecture: Server + Client

  Client A                   Server                    Client B
  ┌──────────┐              ┌──────────────┐          ┌──────────┐
  │CC sessions│              │   Shared DB   │          │CC sessions│
  │    ↓      │  POST        │              │  POST    │    ↓      │
  │ Distill   │──/ingest───→│  Embed+Store  │←/ingest──│ Distill   │
  │ (local)   │              │      ↓       │          │ (local)   │
  │           │  MCP         │ Consolidate  │   MCP    │           │
  │    CC    ←│──(search)───│ (score,link, │──(search)→│   CC     │
  │           │              │  reflect)    │          │           │
  └──────────┘              └──────────────┘          └──────────┘

Server mode — local DB + MCP server + nightly consolidation. Client mode — distill locally for privacy, POST memories to a shared server.

Open source + commercial Pro

Hicortex is MIT-licensed and free forever. The npm package is the complete client: capture, distillation, retrieval, MCP tools, multi-client architecture.

Commercial Pro features (lesson selection engine, validation, cross-agent learning, prescriptive distillation, smart context assembly) are sold separately by Gamaze. Pro is server-side intelligence — no separate npm package, no client-side license keys to bypass. You point your client at a Pro server and the same code calls Pro endpoints if available.

This is the open-core model:

OSS (this repo): the memory client. Anyone can self-host, fork, modify, ship in their own product.
Pro (commercial): the intelligence layer. Funds OSS development, runs as a SaaS or licensed self-host.

See hicortex.gamaze.com for pricing and Pro features.

Project layout

packages/openclaw-plugin/    The npm package (@gamaze/hicortex)
  src/                       TypeScript source
    cli.ts                   CLI entry: server, init, nightly, status, uninstall
    init.ts                  Interactive setup wizard
    mcp-server.ts            HTTP/SSE MCP server (persistent daemon)
    nightly.ts               Nightly pipeline: distill + consolidate + inject
    consolidate.ts           Importance scoring, reflection, linking, decay
    distiller.ts             Transcript → LLM → memories
    storage.ts, db.ts        SQLite + sqlite-vec + FTS5
    retrieval.ts             BM25 + vector search with RRF fusion
    embedder.ts              Local ONNX embeddings
    llm.ts                   Multi-provider LLM client
    features.ts              Centralized feature gating
    claude-md.ts             CLAUDE.md lesson injection
    prompts.ts               LLM prompt templates
    license.ts               License validation
    transcript-reader.ts     Claude Code .jsonl reader
    index.ts                 OpenClaw plugin entry
  skills/                    Bundled OpenClaw skills (/learn, etc.)
  openclaw.plugin.json       OpenClaw plugin manifest

Development

git clone https://github.com/gamaze-labs/hicortex.git
cd hicortex/packages/openclaw-plugin
npm install
npm run build
npm test

See CONTRIBUTING.md for the contribution guide.

License

MIT — see LICENSE.

@gamaze/hicortex

Quick Install