Give your coding agent a long-term memory.

Agent Wiki is one markdown vault that your agents build as they work. They save the decisions, fixes and research they find, and search the vault before they go to the web. Each session starts with what every earlier session learned, in every project and on every machine.

pip install agent-wiki-kb

Free and open source (GPL-3.0). Python 3.10 or later. Works with Claude Code, pi and OpenCode.

Your agent checks the wiki first Example
You

The payments API is timing out under load again. Where do we start?

  1. $ awiki search "payments timeout"
  2. ## Pin pgbouncer to transaction mode
  3. decisions/pgbouncer-mode.md
  4. ## Payments Service API
  5. projects/payments-service-api.md
  6. $ awiki show decisions/pgbouncer-mode.md --section "Why"
  7. Session mode held one server connection per client. Under load the pool ran out and requests waited 30 s.

The agent starts from the decision your team already made, reads only the part it needs, and goes straight to the fix.

How it remembers

You set it up once. After that, the work your agents do goes into the vault, and the vault comes back to them when it is relevant.

The loop
  1. 1
    Sessions file themselves

    Each time an agent starts, a background sync puts every finished session from Claude Code, pi and OpenCode into the vault.

  2. 2
    The agent searches before the web

    awiki search finds the pages that match all your words. awiki show reads a page whole, or one section of it.

  3. 3
    It saves what it learns

    A decision, a fix that was hard to find, a pattern you can use again. The awiki-save skill writes it as a page for the next session.

One command teaches your agent

awiki guide prints a short block for CLAUDE.md or AGENTS.md. Tell your agent to run it, and the agent adds the block to the project itself.

One vault for all your projects

A fix you found in one repo is there for the next repo. The vault is not per project.

Local, or shared over the network

awiki serve shares the vault over HTTP. Other machines use the same awiki commands. Each user gets per-vault roles and tokens, and there is a read-only web viewer.

Safe for smaller models

The CLI is the only way in, so the rules hold for every agent. Sources in raw/ do not change. A page that was edited by hand shows a diff, and awiki does not overwrite it.

It is just markdown files

No database and no server you must run. Each page is a markdown file with YAML frontmatter and [[wikilinks]]. You can grep it, open it in Obsidian or Logseq, or script it in Python.

projects/payments-service-api.md
---
title: Payments Service API
topic: projects
tags: [payments, api, backend]
created: 2026-04-14
updated: 2026-09-02
sources: [raw/payments-service-notes.md]
---

# Payments Service API

REST API for the mobile app and external integrations.
Timeouts under load come from the connection pool, not the
handlers. See [[Postgres connection pooling]] and
[[Pin pgbouncer to transaction mode]].

Get started

Install the CLI, make a vault, and connect your agent. Then use your agent as usual.

Terminal
pip install agent-wiki-kb
awiki init ~/vaults/agent-wiki
awiki skills install
awiki hook install --agent claude
Or let your agent do the setup. Point it at the repo and ask:

Set up agent-wiki from https://github.com/TacoTakumi/agent-wiki on this machine.

Not on Claude Code?

Use --agent pi or --agent opencode. For other tools, --agent manual prints the steps to connect them by hand.

Then start your agent and tell it:

We use awiki here. Run awiki guide and wire it into this project's memory file.

Pairs with specflo

specflo is our spec-driven workflow for coding agents. It takes a feature through brainstorm, spec, plan and execute, with each step kept as markdown on disk. It uses Agent Wiki as its research memory, so each new feature starts from what the last one learned.

Where specflo uses the wiki
  • Brainstorm specflo-brainstorm

    Checks the wiki for earlier decisions and research before it starts on a new idea.

  • Research specflo-research

    Starts from the wiki, checks it against the web, and saves the findings back with their sources and dates.