MemoryAtlas

Engram

Gentleman Programming

Agent-agnostic persistent memory for AI coding agents: a single, dependency-free Go binary backed by SQLite + FTS5 full-text search, exposed via an MCP (stdio) server, an HTTP API, a CLI, and an interactive TUI. Works with any MCP client (Claude Code, Codex, Gemini CLI, Cursor, Windsurf, VS Code Copilot, OpenCode, and more).

Storage
A single local SQLite database (~/.engram/engram.db) with FTS5 full-text indexing — one binary, one file, no Node/Python/Docker dependencies. Designed so memory survives agent context compaction across sessions.
Retrieval
FTS5 full-text search exposed as MCP tools (and via HTTP/CLI/TUI); the agent writes and recalls memories over stdio. 'Engram Cloud' is an optional, self-hosted sync + dashboard layer (Docker/GHCR image you run yourself) for project-scoped replication across machines — local SQLite stays authoritative, not a vendor-hosted SaaS.
Self-host
Self-host: trivial
License
MIT
Pricing
Fully open-source (MIT), free to self-host end to end: the core binary (brew install) and the optional 'Engram Cloud' sync/dashboard layer (self-hosted via Docker Compose or the ghcr.io/gentleman-programming/engram image) are both free — there is no vendor-hosted managed tier or paid plan. · Free / OSS
GitHub stars
5,947
Last release
2026-07-20
Last commit
2026-07-30
First catalogued
2026-06-28

Strengths

  • Zero-dependency single Go binary over one SQLite file — trivial to self-host
  • Agent-agnostic via MCP: works across Claude Code, Codex, Gemini CLI, Cursor, Windsurf, VS Code, OpenCode and others; `engram setup gemini-cli` is a distinct one-liner from `engram setup antigravity-cli` (different config paths — `~/.gemini/settings.json` vs `~/.gemini/config/mcp_config.json` — both still documented as of 2026-07-15)
  • Explicit focus on surviving context compaction; CLI/HTTP/TUI surfaces in addition to MCP
  • No embedding model inference at query time — FTS5 is pure SQLite I/O, making retrieval latency predictable with no ML inference overhead

Watch out

  • 'Engram Cloud' is a self-hosted sync/dashboard extra you run yourself (Docker Compose or ghcr.io image) — not a vendor-managed SaaS with its own pricing; there is no paid tier at all as of 2026-07-23
  • MCP transport is stdio-only (no network MCP endpoint) and the HTTP API binds to loopback only — relevant for Dockerized/remote agents
  • Retrieval is full-text (FTS5) rather than semantic/embedding — note for comparability against vector-based systems
  • Not a Python (or Node) package despite sharing a catalog family with pip-installable libraries: distribution is a single Go binary via Homebrew or `go install`, and the README states outright 'No Node.js, no Python, no Docker.' There is no PyPI wheel to `pip install`; a separate npm package (`gentle-engram`, latest pi-v0.1.9 released 2026-06-30) is a native memory provider for the unrelated 'Pi' agent tool, not a Python or general embedding SDK.

Best for

  • Coding agents needing a lightweight, local, agent-agnostic persistent memory that survives session and compaction boundaries

How it integrates

Benchmark results

No sourced results yet.

Sources

Last verified 2026-08-01 · updated by refresh-framework-cards