Recall Meridian

Local memory layer for AI agents. Works with Claude, ChatGPT, Cursor, Codex, and any MCP client.

Recall Meridian stores per-project knowledge, milestones, decisions, and handoffs in a local knowledge base so your AI agent can pick up where the last session left off — without paying for a chat-history product or trusting a vendor's memory feature.

Two ways in

1. Try the pattern (no install, 30 seconds). Drop RECALL-PATTERN.md into Claude Code as your project's CLAUDE.md. You immediately get: cross-session project memory with a real schema, six typed relationship types (supersedes / contradicts / confirms / qualifies / deprecates / child_of), source citations, and a markdown vault Claude reads on every session. No servers, no SaaS, no installs.

A pattern is not a runtime. Schema enforcement, decay, signed provenance, audit, and search all rely on the engine. The pattern's job is to teach the discipline and produce a vault that imports cleanly when you upgrade.

2. Install the engine (npm, full features). Add typed-graph SQL queries, signed provenance, an automated holdout-gated self-improvement loop, a 17-module security stack, scheduled defenses (DLP, canaries, dream cycle, adversary engine, anchor snapshots), and multi-machine sync:

npm install -g @jesseneff1/recall

Validate the work you already did, then import it (zero data loss):

recall pattern-validate ./        # health score 0-100, finds drift
recall import-vault ./ --repair   # imports, auto-fixes recoverable issues

Same schema, both modes. Upgrade is additive — the pattern is real on its own, and the engine adds on top of work you already did. See RECALL-PATTERN.md for the full comparison table.

Initialize a project:

mkdir my-thing && cd my-thing
recall init my-thing

First useful workflow

# Capture a decision so future-you doesn't re-litigate it
recall add my-thing decisions "switched to postgres" --because "row-level locking"

# List what you've captured
recall query my-thing "TABLE name FROM *"

# Get a project handoff (the magic moment)
recall my-thing

The last command prints a ~300-word brief any AI assistant can ingest: milestones, open TODOs, recent decisions, next step. Paste it into Claude, ChatGPT, Cursor — anywhere.

What this project is

Recall Meridian is not a chat memory feature, not a notes app, not an IDE rule file, not a vector database.

Core concepts

Three layers in one stack:

| Layer | Size | What | |---|---|---| | L1 — Pointer | ~50 tokens | A breadcrumb your AI can follow | | L2 — Summary | ~3 KB | Project state in narrative form | | L3 — Full JSON | On demand | Every entry, every relationship |

Most "memory" tools dump everything into context and hope for the best. Recall Meridian starts with a 50-token pointer, expands to a 3 KB summary when the agent asks, and only loads full JSON when you specifically need it.

How it's different

| | Slash commands / chat memory | Recall Meridian | |---|---|---| | Storage | Vendor-bound | Local SQLite | | Scope | One chat / one IDE | Cross-tool, cross-session | | Audit | Opaque | Full history, exportable | | Trust model | Whatever the model remembered | Draft vs trusted entries with promotion gates | | Lock-in | High | Apache 2.0, your data on your disk |

Slash commands

Twelve commands installed. Five most-used:

| Command | Purpose | |---|---| | /recall | Project handoff — milestones, TODOs, next step | | /recall-sync | Pull latest session data into the dashboard | | /recall-kb | Add/list/update KB entries by category | | /recall-milestone | Mark milestone complete or queue next | | /recall-analyze | Run AI analysis over recent sessions |

Full list: recall --help.

Works with

• Claude Code (CLI + IDE)
• Claude.ai web (via MCP)
• ChatGPT (via MCP)
• Cursor (via MCP)
• Codex CLI
• Any MCP-compatible client

Architecture

Hexagonal: a stable core of invariants, a contract layer features must declare against, and pluggable adapters. Every feature run passes through draft → reviewed → trusted with an append-only audit ledger.

Storage is local SQLite + FTS5 search under ~/.recall/. No embeddings server required; optional semantic search via your own OpenAI key.

Full architecture map, file-by-file inventory, and "how to plug in a custom feature": HEXAGONAL.md.

Docs

• docs/setup/getting-started.md — first-run walkthrough
• docs/setup/feature-registry.md — feature lifecycle
• AGENTS.md — instructions for AI agents working in this repo
• CONTRIBUTING.md — how to contribute

Status

Pre-1.0. The CLI works today. Desktop app and the Meridian publishing protocol are in development — join the waitlist.

Local-first by design. Your knowledge base never leaves your machine unless you publish to a Meridian server (opt-in, future).

What this is not ready for

• Multi-user shared memory. Per-machine local KB only — no team server yet.
• Production secrets vault. Use a real secrets manager.
• Autonomous publication. Anything that escapes the local KB requires explicit promotion.
• Replacement for source control. It augments git history; it doesn't replace it.

License

Apache 2.0. See LICENSE.