Shared fleet memory for AI agents: nightly self-correction, recall every prompt (supported agents).
io.github.gamaze-labs/hicortex MCP Server
The MCP server io.github.gamaze-labs/hicortex provides “Shared fleet memory for AI agents.” Its scope includes nightly self-correction and “recall every prompt,” with support explicitly described as applying to supported agents.
🛠️ Key Features
Shared fleet memory for AI agents
Nightly self-correction
Recall every prompt (for supported agents)
🚀 Use Cases
Persisting and reusing context across AI agents within a fleet
Improving agent behavior via recurring nightly self-correction
Enabling prompt-level recall for supported agents
⚡ Developer Benefits
Consistent shared memory semantics for AI agents
Prompt recall capability to support agent reasoning continuity
Explicit support model: recall applies to supported agents
⚠️ Limitations
Support is limited to “supported agents” as stated
No additional tools, endpoints, or configuration details are provided in the available source data
Shared memory for AI agents — it corrects itself overnight, and what one agent learns, the whole fleet knows. One memory across every agent, every project, every machine — they stop assuming and start knowing.
One brain, every harness — Claude Code, Hermes, OpenClaw, Pi, OpenCode, and any MCP-compatible agent share the same memory.
Pushed, not pulled — in every supported coding agent, a compact recall index is injected on every prompt, so the decisions, corrections, and context an agent needs are already in front of it. No re-explaining, no copy-paste, nothing to maintain. Zero LLM calls per turn — no API cost or rate-limit hit from recall.
Consolidates overnight — each night it reads the day's sessions, distills what matters, and turns it into Learnings, links, and a knowledge graph.
Local-first — raw sessions never leave the machine; only distilled memory is stored.
Install
bash
npx @gamaze/hicortex init
Claude Desktop: one "yes" during init.
Auto-detects your environment, configures one LLM (Ollama, the Claude CLI, or an API key), installs a local daemon (launchd on macOS, systemd on Linux), and registers MCP tools with Claude Code.
For multi-machine setups, point thin clients at a shared server — no local DB or LLM on the clients:
init auto-detects the other harnesses and installs their clients: a Pi extension (~/.pi/agent/extensions/hicortex.ts — pushed recall, identity + lessons, the ten tools; or copy pi-extension/hicortex/index.ts there manually), an OpenCode plugin (~/.config/opencode/plugins/hicortex.ts — the same trio; or copy opencode-plugin/hicortex/index.ts there manually), the Hermes plugin, and the OpenClaw plugin. pi-mcp-adapter remains a generic MCP escape hatch for any harness (verified against the SSE endpoint) — Pi no longer needs it. See the install docs.
How it works
code
CAPTURE (nightly) CONSOLIDATE (nightly) RECALL (every prompt)
sessions → denoise score · reflect · link a compact index of
→ POST /distill decay · dedup · supersede relevant memories is
(one model, all phases) pushed into the prompt
→ full text lazy-loaded
Memories strengthen when agents use them, fade when they don't, and link to related ones automatically. Retrieval is hybrid BM25 + vector search — zero-LLM at query time.
The first nightly run captures the last 7 days of sessions by default (not your entire history) — run hicortex nightly --recapture-window <days> once to import more.
Features
Per-prompt recall push — relevant memory lands in context every turn; the agent fetches full content with hicortex_get only when it needs it.
Memory analytics at /dashboard — growth, recall adoption, and a nightly digest of what was learned.
Knowledge graph at /viz — memories clustered by domain, connected by relationship edges.
Domains & tags — multi-tag classification with a configurable vocabulary; your categories drift with your data.
Learnings from reflection — nightly reflection extracts general, reusable Learnings, not just Experience logs.
Self-correcting store — every night, stale facts are rewritten in place with dated provenance; near-duplicates resolve into one (verbatim copies kept free, merges recoverable); superseded decisions are demoted, never re-surfaced. No zombie memory.
Unprompted by design — coding agents get recall injected via hooks; instruction-capable clients (Claude Desktop, Cursor-class) get standing instructions, so memory is used without being asked. Plain MCP clients keep full search.
Self-calibrating — recall, decay and merge boundaries report their own statistics; tuning is measured, never guessed.
Standing context layer — hand-edited "who you are / how to work" Markdown, injected every session, never decayed.
MCP
Nine MCP tools — hicortex_search, hicortex_get, hicortex_recent, hicortex_ingest, hicortex_lessons, hicortex_index, hicortex_graph, hicortex_update, hicortex_delete — plus a /learn skill to save explicit learnings. Full reference →
Stack
TypeScript · Node.js 20+ · SQLite + sqlite-vec + FTS5 (semantic + full-text in one DB) · ONNX embeddings (bge-small-en, CPU) · MCP over HTTP/SSE · one configurable LLM (Ollama, Claude CLI, or any OpenAI-compatible endpoint).
Development
bash
git clone https://github.com/gamaze-labs/hicortex.git
cd hicortex
AGENTS.md at the repository root defines the machine-checkable verification contract. "Done" means the full command chain exits with code 0. The contract mirrors what CI runs. Contributors — human or agent — run it before claiming work complete.
