Open-source context infrastructure for AI coding agents
Stop explaining your codebase
to AI every new session.
Recall scans your repository and generates persistent, evidence-backed context that coding agents can reuse across sessions.
npx recall-context initWorks with Claude Code, OpenAI Codex, Gemini CLI, Cursor.
Every new session starts here.
Most coding agents only know what you paste into the current conversation.
- Architecture decisions.
- Naming conventions.
- Important dependencies.
- Project history.
- Previous trade-offs.
They disappear.
So every new session starts from zero.
Without Recall
Developer:
How does authentication work?
Agent:
I do not know this repository yet.
Agent:
Can you explain the architecture, conventions, and relevant files?
With Recall
- Architecture
- Conventions
- Decisions
- Risks
- Technical debt
Context ready.
Recall does not replace the coding agent.
It gives the agent the repository context it did not have.
How Recall works
Recall analyzes your repository once, builds a structured project memory, and gives every future AI session immediate access to the parts of the codebase that matter.
Scan Repository
Analyze project structure, frameworks, dependencies and conventions.
Build Memory
Generate architecture, decisions, conventions, risks and technical debt.
Understand Task
Rank the most relevant files for the current request.
Provide Context
Produce deterministic context that any coding agent can immediately use.
CLAUDE.md and AGENTS.md vs. Recall
Recall doesn't replace hand-authored instruction files — it complements them with generated, evidence-backed context.
| CLAUDE.md / AGENTS.md | Recall | |
|---|---|---|
| Authored by | A human, by hand | Derived from repository evidence |
| Stays current? | Can go stale silently | Detects and reports staleness (`recall status`) |
| Shape | Typically one static document | Snapshot + multiple focused files |
| Scans the repo? | No | Yes — deterministic, local scan |
| Evidence paths | Not by default | Every claim traces back to source |
What Recall generates
- Deterministic, local-first
- Every run is a local, deterministic scan. No account, sign-up, or API key required.
- Evidence-backed
- Findings carry file paths as evidence — nothing is asserted without a source.
- Explainable ranking
- Task-focused context ranking exposes its reasons in JSON, not just a black-box score.
- Staleness detection
- `recall status` and `recall doctor` report when memory has drifted from code.
- Human content preserved
- Generated Markdown wraps Recall’s content in markers; anything you write outside them survives every update.
- No telemetry
- Recall does not call home. Everything runs on your machine, in your repository.
Try it in your own repository
npx recall-context initrecall context --task "Fix password reset controller" --stdoutWorks with any agent that reads text
Recall's output is plain Markdown and JSON, so it works with any coding agent that can read a file or stdin — including:
- Claude Code
- Codex
- Gemini CLI
- Cursor
- Any agent that can read Markdown or JSON
Open source, MIT licensed
Recall is developed in the open. Read the source, file an issue, or contribute on GitHub.
View source on GitHub