Recallant

Self-hosted governed memory for Codex and MCP AI agents.

Recallant gives AI coding agents a durable memory layer for real project work. It records decisions, checkpoints, evidence, reviewable rules, and source-backed context so a later session can resume without asking maintainers to rebuild the same background over and over.

Recallant is private by default, Apache-2.0 licensed, and designed for OSS maintainers who use Codex, Cursor, Claude Code, Windsurf, or any MCP-compatible client.

Why It Matters

AI-assisted development often loses the thread between sessions. Project decisions live in chat history, local notes, PR comments, terminal output, and human memory. Generic logs or basic RAG can recover fragments, but they usually miss authority, provenance, review state, and scope.

Recallant is built around governed memory:

• Evidence first: raw workflow evidence, source references, and bounded excerpts stay attached to remembered facts.
• Rules need authority: durable guidance cannot silently become an instruction just because an agent inferred it.
• Project context stays scoped: current-project memory is the default; cross-project examples are explicit and labeled.
• Maintainers keep control: destructive actions, paid API use, public exposure, secrets, and global rules stay behind policy gates.

The result is a memory system that helps agents work faster without turning old context into an unreviewed pile of instructions.

What You Can Try Today

Recallant is pre-release, but the first coding-agent memory slice is working: attach a project, start an agent-backed session, read a context pack, write decisions/actions/tests/checkpoints, close out, and recall that memory in a later session.

Install:

curl -fsSL https://raw.githubusercontent.com/Mushkrot/Recallant/main/scripts/install-recallant-bootstrap.sh | bash

Onboard a project for Codex:

cd /path/to/project
recallant onboard --client codex --install-local-hooks --verify

Or run the step-by-step proof path:

recallant attach .
recallant connect codex --project-dir . --dry-run
recallant connect codex --project-dir .
recallant demo-capture --project-dir .
recallant doctor --project-dir . --require-capture
recallant ask "what did the agent remember?" --project-dir .

The important state is not "installed" or "configured". The important state is capture active: Recallant has observed real context reads, memory writes, and checkpoints for the project.

Product Shape

Recallant runs as a local or self-hosted memory service:

• CLI for install, attach, connect, doctor, capture proof, and cleanup.
• MCP server for agent clients.
• Postgres/pgvector-backed storage.
• Private Workbench UI for review, rules, source context, settings, and management chat.
• Local-first model routing with explicit approval gates for paid APIs.

See Architecture for the public system overview.

Documentation

• Quickstart: install Recallant and prove one project can remember.
• Why Recallant: problem statement and community value.
• Comparison: inspirations, alternatives, and the gap Recallant fills.
• Self-hosting: profiles, rollback, verification, and security defaults.
• Client setup: Codex and other MCP clients.
• Security: public threat model and safe defaults.
• Roadmap: pre-release status and next milestones.
• Contributing: how to work on the project.

Status

Recallant is pre-release. It is suitable for local evaluation and development, not for unreviewed team-wide production rollout.

Current strengths:

• first end-to-end coding-agent memory loop;
• Codex-first MCP workflow with generic MCP client posture;
• private-by-default Workbench and server defaults;
• explicit security and cost governance design;
• smoke coverage for the core capture/recall path.

Known pre-release work:

• final clean-host install validation;
• more public screenshots and docs polish;
• broader client pilot matrix;
• security review and release hardening;
• packaging and versioned release tags.

License

Apache License 2.0. See LICENSE.