Session memory for Claude Code

Your AI should remember things.

agent-recall gives Claude Code a filing cabinet instead of a goldfish brain. Save context, distill it, search it later, and resume without replaying your own work back to the model.

open source · MIT · Python 3.8+
/install agent-recall
circle the useful part, toss the rest

Session memory for Claude Code.

Save context, distill it, resume without starting over.

No more ten-minute re-briefs.

What gets kept?

  • decisions
  • files touched
  • recurring failures
  • what to do next
The Index

What's in the box

Three commands, one memory system. The cards are separate on purpose: save sessions, learn from failures, and inspect history without dragging a whole framework into the room.

file it where future-you can find it

/recall

Search sessions, save restart prompts, and pick work back up mid-thought.

session lookup
$ /recall auth

2 matches
2026-03-14 switched JWT to RS256
callback URL had to match env exactly

/failures

Track the commands that keep hurting you and surface the fix while the error is still warm.

failure SOP
$ git push origin feature/auth
fatal: permission denied

recall SOP: switch HTTPS to SSH
you hit this three times this month

/history

Zero-hook command history when you just need to retrace what you actually ran.

command trail
$ /history --failures

python3 sync.py
rg "clinicSelected"
pytest tests/test_share.py
How It Works

Save. Distill. Resume.

The point is not preserving the transcript. The point is preserving the state of the work without carrying every tangent, typo, and dead-end into the next session.

Save

/recall save captures messages, commands, failures, and current state before the session expires.

step 1
$ /recall save
47 messages
12 commands
2 failures

Distill

Noise comes out. Decisions, files changed, risks, and next steps stay in the briefing.

step 2
kept: branch state, blockers
cut: dead-end retries
wrote restarts/payroll-fix.prompt

Resume

Launch the saved restart prompt and continue from the middle instead of the beginning.

step 3
$ /recall restart 1
branch: feature/payroll-fix
next: verify rounding edge case
Platforms

Works wherever you code

agent-recall is one memory index with a few different doors into it. The shape changes per tool; the session history does not.

Claude Code

  • SessionStart hook
  • auto-loads recent context
  • native slash commands
$ /install agent-recall

Codex

  • AGENTS.md integration
  • shared index with Claude Code
  • same restart prompts
python3 ~/.claude/bin/session-context.py "$PWD"

Gemini CLI

  • GEMINI.md integration
  • project memory handoff
  • shared command history
reads the same recall index

Claude Desktop

  • filesystem MCP bridge
  • browse saved sessions
  • project-level history
point MCP at ~/.claude/projects
The agent-* Family

The toolkit

Three small tools with the same visual language and the same bias: stop dragging context around by hand.

agent-recall

Session memory, failure recall, restart prompts, and searchable work history.

Open the repo ->

agent-pb

Prompt builder and context assembly for shaping the right input before the run starts.

Visit agent-pb ->

agent-look

Visual capture and browsing help for when the work lives in screenshots, pages, or interfaces.

Visit agent-look ->
Install

Put it in the drawer

Primary path is the plugin marketplace. Manual clone stays as the fallback if you want to inspect the wiring yourself.

# Claude Code (plugin)
$ /plugin marketplace add ashrocket/agent-recall
$ /install agent-recall

# Or clone and configure manually
$ git clone https://github.com/ashrocket/agent-recall
$ cd agent-recall
# follow README for settings.json hook config