Remembra Relay
One coding agent stops, the next one already knows: session handoffs built from git facts, across machines and vendors.
- Source repo
- remembra-ai/remembra
- Stars
- ★ 17
- Last updated
- 10d ago
- License
- MIT
- Primary language
- Python
- FA score
- 83/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformClaude CodeCodex (Partial support)
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A solo developer switching the same repo between a laptop, a server and worktrees who wants each new session to start with real context
- Not a fit if
- Teams unwilling to send session data to Remembra Cloud (unless self-hosting)
- Coding workflows that do not use git
- Users relying on Cursor, Gemini CLI or other unverified hooks
- Website
- remembra.dev ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-11latest release v0.17.8 (2026-11-11)
- Source review
- 83/100 · Good
What does this agent do, and when should you use it?
Remembra Relay is an MCP server plus CLI (remembra-relay) that produces session handoffs built from git facts when a coding agent's session ends. On close, it reads the branch, commits, changed files and unpushed commits from git, and for Claude Code and Codex also test runs from the local transcript; on the next session, the brief command or the session_brief MCP tool delivers what is done, failing and next in about 1,500 tokens, checking the agent's own summary against the recorded facts. Projects are identified by git remote, so handoffs work across machines, worktrees and vendors, and every handoff is kept in order in a trail. Underneath is a standalone memory API (Python/TypeScript SDKs, REST API, 21 MCP tools) with fact extraction, entity resolution, hybrid retrieval and PII protection. It runs on Remembra Cloud (free tier available) or self-hosted under the MIT license with Docker, Qdrant and Ollama.
At session end, remembra-relay close reads the branch, commits, changed and uncommitted files and unpushed commits from git; for Claude Code and Codex it also parses test runs and open items from the local transcript (which never leaves the machine, with secrets redacted before anything is sent). At the next session start, remembra-relay brief or the session_brief MCP tool renders a ~1,500-token brief: who worked last, what is done, what is failing and the next step — the done/not done/failing sections come from git and test runs, never from an LLM, and the agent's own summary is shown as unverified or contradicted against those facts. remembra-relay trail lists every handoff in order. The MCP server (remembra-mcp) exposes 21 tools: store_memory/recall_memories for memory, search_entities/timeline for entities and time, send_to_inbox/get_inbox for messaging between agents, and more. The underlying memory layer is directly usable via the Python SDK (from remembra import Memory; memory.store(...); memory.recall(...)) or the REST API.
- A solo developer switching the same repo between a laptop, a server and worktrees who wants each new session to start with real context
- A developer who uses Claude Code by day and Codex CLI by night and needs vendor-agnostic handoffs
- An engineer picking up someone else's work who wants ground truth from git rather than the previous agent's self-report
- Users of Cursor, Gemini CLI or any MCP client, who read briefs and write handoffs through MCP tools
- Developers adding a persistent memory layer with entity resolution and hybrid retrieval to an app via the Python or TypeScript SDK
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pipx install --force 'remembra[mcp]>=0.16'curl -sSL https://raw.githubusercontent.com/remembra-ai/remembra/main/quickstart.sh | bash
- Config files
docker-compose.quickstart.yml- Install notes cover
- Docker
How do you install or deploy this agent?
Cloud route (get a free key at app.remembra.dev first):
bash
pipx install --force 'remembra[mcp]>=0.16'
remembra-install --all
remembra-relay connect --applySelf-hosted (starts Remembra + Qdrant + Ollama locally, no API keys needed):
bash
curl -sSL https://raw.githubusercontent.com/remembra-ai/remembra/main/quickstart.sh | bashOr with Docker Compose:
bash
git clone https://github.com/remembra-ai/remembra && cd remembra
docker compose -f docker-compose.quickstart.yml up -dHow do you use this agent?
After installing, run remembra-relay connect alone first to preview every hook change without writing, then connect --apply to write them (each file is backed up); in Codex, run /hooks once to trust the hooks. Use your coding agent as usual: the close hook fires at session end, and the brief hook or the session_brief MCP tool delivers the handoff at session start. Review handoffs manually:
bash
remembra-relay trailVerify setup:
bash
remembra-doctor allUse the memory API directly:
python
from remembra import Memorymemory = Memory(user_id="user_123")
memory.store("Had a meeting with Sarah from Acme Corp. She prefers email over Slack.")
result = memory.recall("How should I contact Sarah?")
print(result.context)What are this agent's strengths and limitations?
- Handoff facts come from git and test runs, not an LLM; the agent's own summary is checked and the README shows a concrete example of it being contradicted
- Projects are identified by git remote, so handoffs work across machines, worktrees and vendors (Claude Code, Codex, and any MCP client)
- Fully self-hostable under MIT (Docker + Qdrant + Ollama), free of charge
- Includes a complete memory layer: 21 MCP tools, Python/TypeScript SDKs, entity resolution, timeline queries and an inbox between agents
- Core handoff depends on Remembra Cloud (free tier exists, but a key is required); full offline use requires deploying your own server
- Only Claude Code and Codex session hooks are verified; hooks for Cursor, Gemini CLI, Qwen Code and Kimi are unverified and only installed with --include-unverified
- Codex support is verified against a prerelease, codex-cli 0.155.0-alpha.16.4; no stable Codex release has been tested
- Benchmark evidence is thin: 1 of 10 LoCoMo conversations (76%, all 47 adversarial questions scored 0); a full run has not been done
How does this agent compare with similar options?
The README compares Relay to three alternatives (docs/comparisons/handoff-tools.md, with sources and dates): vendor memory (Claude Code, Codex, Copilot, Windsurf) is not cross-vendor and keeps no durable trail; local handoff tools (such as claude-mem and agentmemory) do not work across machines and rely on transcripts/diffs; memory APIs (Mem0, Zep, Letta) work across machines but their handoff facts do not come from git. Relay's differentiators are cross-vendor support, git-fact verification and a durable trail.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Remembra Relay This agent | 83 · Good | Agent plugin / skillFreemium | ★ 17 | 10d ago | Python | Claude Code |
| Happier | 62 · Some gaps | Desktop appFree + model costs | ★ 2.2k | today | TypeScript | Codex · Claude Code |
| Emulo | 76 · Good | Agent plugin / skillFree + model costs | ★ 306 | 1mo ago | HTML | Codex · Claude Code |
| DevSpace | 64 · Some gaps | CLIFree + model costs | ★ 5.5k | 19d ago | TypeScript | ChatGPT |
How does FollowAgents rate this agent?
Why each dimension lost points
The installer is documented to show changes and ask before writing; Relay can preview changes and keeps backups. Keys are read through a hidden prompt or environment variable. The docs say transcripts stay local, recognized secrets are redacted from sent data, and SQLite memory text remains plaintext. These controls are substantial, but redaction is pattern-based, encryption covers limited fields, and the supplied material does not fully show write permissions and confirmation boundaries for CLI and MCP operations. Attribution is only partly supported by package author/maintainer fields and a security contact; publisher identity remains unverified.
Relay CLI documentation describes time budgets, timeout fallback, queueing and retry for failed delivery; doctor provides read-only diagnostics with targeted fixes. README, configuration and installation guidance broadly agree, but the supplied material has version differences (package metadata says 0.17.6 while many examples use 0.16.x), and some support claims are tied to specific client versions. These weaknesses reduce self-consistency and dependency availability scores.
The materials clearly target cross-agent coding-session handoffs and collaboration, describe MCP tools and client support status, and distinguish git/transcript-collected facts from agent-declared unverified facts and unverified client integrations. Hooks, MCP, CLI, self-hosting and multiple client configurations cover varied environments, with scenarios and limits stated clearly.
README, installation, configuration, security, troubleshooting and Relay guidance are clearly organized. Installation, diagnostics, examples, troubleshooting, limitations and the MIT license are well covered. Naming and release/version information exist, but package version, install floor and example versions do not fully align, and the supplied packet does not include the full changelog. Maintainer fields and a vulnerability-reporting channel are present, while maintenance responsibility and update commitments are only partly clear; publisher identity is unknown.
Relay turns session facts from git and some client transcripts into a concise handoff, next steps and a trail, while flagging contradictions and unverified agent claims. This has a clear cross-machine, cross-vendor use. Its value depends on a Remembra service, keys, hook installation and varying client support, which add setup cost.
The README gives concrete descriptions of core workflows, fact sources and verification states; CLI source documentation adds budgets, queue behavior, hooks and project resolution. Security, configuration and installation documents offer partial corroboration. Much product behavior has no corresponding implementation body in the supplied source excerpts, and only part of CI is provided, so corroboration is incomplete. The materials explicitly separate collected facts, inference, agent declarations and unverified status.
- Publisher identity is unverified; package author and maintainer metadata do not establish verified organizational identity.
- Handoffs can contain sensitive context. Secret redaction relies on known patterns, and the security docs state that SQLite memory text and some indexes are not covered by field encryption.
- Installation writes agent configuration and hooks. Review the preview first; hook verification and supported versions vary by agent.
- Check version information: package metadata says 0.17.6, while supplied install commands and security release examples mostly use 0.16.x.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * Release 0.17.8: restore recovery, queued waits and hook diagnostics by @Mani87-nq in https://github.com/remembra-ai/remembra/pull/91 Full Changelog: https://github.com/remembra-ai/remembra/compare/v0.17.7...v0.17.8View on GitHub ↗ -
Release notes excerpt
What's Changed * 0.17.7: fixes from the Codex/Claude fault survey (49 faults, frozen paths untouched) by @Mani87-nq in https://github.com/remembra-ai/remembra/pull/90 Full Changelog: https://github.com/remembra-ai/remembra/compare/v0.17.6...v0.17.7View on GitHub ↗ -
Release notes excerpt
Full Changelog: https://github.com/remembra-ai/remembra/compare/v0.17.5...v0.17.6View on GitHub ↗
FAQ
Is it paid?
Does my code or transcript leave my machine?
I use Cursor or Gemini CLI — does it work?
How much of the brief is LLM-generated?
Will multiple agents in the same repo conflict?
Is Remembra Relay open source?
Is Remembra Relay safe to use?
Is Remembra Relay still maintained?
Who makes Remembra Relay?
What is Remembra Relay's official website?
Thanks — we use this to improve the page.