Khipu — AI skill for Claude Code

Khipu

**Khipu gives AI coding agents a memory that outlives the session.**

The name comes from Andean knotted-cord records (*khipu*): memory encoded as structure, not as a pile of chat transcripts.

What it is

Every coding agent forgets. Close the window and the decisions, the constraints and the approaches you already ruled out are gone.

Khipu is a Mac app, a Python CLI and a set of agent integrations. It captures what happened in an agent session, stores it in PostgreSQL as searchable prose plus a knowledge graph, and puts it back in front of the agent when it is relevant. It works across machines.

flowchart LR
  A["Agent session
ends or compacts"] --> B["Hook queues
a capture job"]
  B --> C["Background drainer
summarises it"]
  C --> D[("PostgreSQL")]
  D --> E["Next session
gets prior work"]
  E --> A

Who it is for

People who use coding agents every day, on one Mac or several, and want the agents to remember. It is a personal system: one database, your machines, your agents.

What it does

• Capture. Hooks in your agents queue a capture after about five user turns or twenty minutes, and always at compaction or session end. A background step summarises each window and writes it to the database.
• Recall. Prior work is pushed into context at session start and on each prompt (Claude Code and Codex). Agents and you can also search by best match, exact words or meaning, or walk the graph.
• Commitments. Open loops from sessions become an Owed list you can close, reopen or snooze.
• Decisions. A registry tracks which decisions still stand, and which were superseded or retracted.
• Briefs and reflect. Optional, off by default: source-backed topic briefs, and answers from memory that cite their sources.
• Libraries. Index a folder of .txt and .md files as another search space.
• Routing and plan usage. Shows each subscrip