Recallant

Self-hosted governed project memory for coding agents.

Recallant lets an agent resume project work from scoped, source-backed context while keeping the maintainer in control of what becomes reusable memory or instruction-grade guidance. It records decisions, checkpoints, evidence, review state, and agent activity in PostgreSQL with pgvector.

The problem is familiar: useful context is scattered across chat history, terminal output, notes, pull requests, and somebody's memory. Saving all of it as a transcript creates a different problem. Old guesses look authoritative, project boundaries blur, and nobody can explain why a rule exists. Recallant keeps evidence, governed memory, and active rules separate.

See It Working

These screenshots come from the real private Workbench with deterministic synthetic data. The generator runs the same browser fixture used for Playwright acceptance testing.

Home shows whether capture and the governed memory loop are actually active.

Review keeps a proposed memory tied to its source and asks for an explicit decision.

Activity connects an observed failure to retry, remediation, and successful verification.

Related MCP server: agentmem

What Ships Today

• One onboarding path. recallant onboard <project> uses reachable local storage or scoped remote access supplied by project configuration, deployment environment, or --server-url. If neither route is configured, it stops before changing project files and shows the setup choices.

• Durable agent lifecycle. MCP and CLI paths cover session start, bounded Context Packs, event capture, governed memory, checkpoints, closeout, and next-session recall.

• Review before authority. Memories retain provenance, scope, status, confidence, and source references. An agent-authored fact does not silently become a binding rule.

• Private Workbench. Home, Ask & Search, Review, Sources, Activity, Settings, and Diagnostics give maintainers a human control surface for memory and capture health.

• Observable agent work. Run replay, grouped errors, recovery chains, capture coverage, and a redacted System Activity Ledger make failures and gaps inspectable without keeping a second raw transcript.

• Private-by-default operations. Local MCP, authenticated scoped remote MCP, backup and restore, explicit cleanup, secret references, and confirmation gates keep sensitive actions server-side.

• Local-first routing. Embeddings can run through a local deterministic or Ollama route. Paid providers require an explicit governed capability and approval.

The Memory Contract

Recallant does not give every stored sentence the same authority. The write path preserves three useful distinctions:

1. Evidence records what happened or what an approved source contained. Agents may inspect it, but it cannot act as a rule.

2. Usable memory holds a reviewed fact, decision, lesson, action, or checkpoint that may help later work inside its declared scope.

3. Active rules contain durable guidance that an authorized human promoted from source-backed memory. Agent inference alone cannot create one.

At session start, the server builds a bounded Context Pack from the current checkpoint, accepted rules, relevant project memories, and recovery state. Deeper evidence stays available through search instead of being pushed into every prompt. Cross-project examples require an explicit query and stay labeled as examples until they are adopted in the current project. This keeps the startup context small and makes the answer to “why does the agent believe this?” inspectable.

Architecture

Recallant has one authority for project identity, memory, provenance, lifecycle, and policy. Agent clients contribute evidence through MCP. The Workbench reads and changes governed state through the same service boundary.

flowchart LR
  Maintainer["Maintainer"] --> Workbench["Private Workbench"]
  CLI["Recallant CLI"] --> MCP["MCP server"]
  Codex["Codex hooks + MCP"] --> MCP
  Other["Other MCP clients"] --> MCP

  MCP --> Policy["Governance and policy"]
  Workbench --> Policy
  MCP --> Audit["System Activity Ledger"]
  Policy --> Context["Context Pack Builder"]

  Policy --> Store[(PostgreSQL + pgvector)]
  Context --> Store
  Audit --> Store

The detailed architecture covers project bootstrap, write and read paths, the governed graph, agent observations, remote access, data lifecycle, and safety boundaries.

Quickstart

For a local single-user evaluation, preview the installer if you want to inspect its plan:

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

Install the local service, then onboard a project:

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

For a pinned prerelease, use the versioned command after the tag is published:

curl -fsSL https://raw.githubusercontent.com/Mushkrot/Recallant/v0.1.0-dev.0/scripts/install-recallant-bootstrap.sh \
  | bash -s -- --ref v0.1.0-dev.0

Tagged installs remain on that prerelease channel during onboarding. Switching to the moving development channel requires an explicit reinstall with --ref main.

The local self-host path may require Docker and PostgreSQL. On a workstation connected to an existing Recallant server, use the same recallant onboard <project> command with that server URL provided by project configuration, deployment environment, or --server-url, rather than installing a second storage stack. See the Quickstart for prerequisites, dry-run behavior, expected readiness proof, rollback, and remote setup.

Onboarding aims to prove more than configuration. A healthy run distinguishes access, context read, semantic recall, memory-loop readiness, and fresh automatic capture. It prints the exact missing proof when one of those states is incomplete.

Current Scope and Limits

Recallant v0.1.0-dev.0 is a development prerelease, not stable production software. It is suitable for local evaluation and controlled development use. Broader project pilots, native cross-client parity, and stable support guarantees are still open work.

Native automatic capture is Codex-first today. Claude Code, Cursor, Windsurf, and generic MCP clients can be configured to use the shared MCP service, but configuration does not prove equivalent native capture. Real bidirectional Codex to Claude Code to Codex continuity, with no pasted human recap, is the next product milestone.

Recallant can attach governed sources and preserve imported evidence now. It does not yet provide a general human document workspace, unrestricted folder crawling, passive Obsidian synchronization, or production-ready chat over arbitrary PDF and DOCX collections. See Current Status and the Roadmap for the maintained boundary.

The four ordered milestones are Cross-Client Continuity, Human Document Memory, a Recallant-owned Task Handoff Record, and then broader human-memory sources and connectors. Shipped graph and observability capabilities remain supported, but further expansion follows those product proofs.

Engineering Evidence

CI runs formatting, lint, TypeScript builds, dependency audits, public security checks, database integration, lifecycle and onboarding acceptance, Workbench smokes, and Playwright browser tests. The browser fixture checks desktop and mobile layouts, required navigation, clipped controls, horizontal overflow, console errors, page errors, and failed requests.

For an ordinary change, start with:

npm run format:check
npm run lint
npm run build
npm run public-readiness:smoke
npm run public-security:smoke

Focused commands and the full contributor workflow are documented in CONTRIBUTING.md.

Documentation

• Evaluate: Quickstart, Current Status, and Product Contract Status.

• Use: Client Setup, Workbench UI, and Agent-Ready Projects.

• Operate: Self-Hosting, Operations Runbook, and Security.

• Understand: Architecture, Domain Language, Agent Observability, and Why Recallant.

• Explore secondary design context: Comparisons and Research Notes.

The full curated route is in the documentation index.

License

Apache License 2.0. See LICENSE.