Goals
Last verified against the code: 2026-10-21 (first commit)
Why Open Memory exists and how far along each goal is. IMPLEMENTATION.md shows how each goal is built, ARCHITECTURE.md how the pieces fit, ROADMAP.md what comes next, and the paper why it is designed this way.
Purpose
Every mistake, made once, everywhere, by morning. An AI agent that keeps what it learns from its own surprises, consolidates it in sleep, keeps only what predicts, and has it in mind at the start of the next session; and a protocol by which what one agent learns, once verified, reaches every agent.
Who it is for
- People who work with a coding agent every day and are tired of repeating themselves and watching it repeat its mistakes. Today: Claude Code users, on their own machine.
- acific.labs itself, whose products are built by agents with no person reviewing code (handbook idea 1): the agents' lessons are the lab's institutional memory.
- Researchers and builders of agent memory, through the paper and the protocol.
Goals
| # | Goal | Status |
|---|---|---|
| 1 | Read the raw stream: every Claude Code session transcript into episodes, idempotently | Done (openmemory ingest) |
| 2 | The unit of memory is the surprise: failed tool calls and the person's corrections become their own episodes | Done (errors tied to the tool call that caused them; corrections detected by opening words) |
| 3 | Recall that weighs meaning, time and use together | Done (0.6 relevance, 0.25 recency, 0.15 strength; hashed embeddings) |
| 4 | Sleep: decay, consolidate surprises into lessons, verify and make predictions | Done (openmemory sleep) |
| 5 | Lessons are observations, never instructions | Done (imperative openings rejected; deterministic texts are descriptive) |
| 6 | Have it in mind, not look it up: lessons rendered where the agent loads them at session start | Done in its weakest form (Claude Code's memory folder); compiled state is goal 12 |
| 7 | The owner sees and vetoes: a morning report; deleting a lesson file is final | Done (morning.md, vetoed.json) |
| 8 | Recall mid-task: an MCP server with recall and remember |
Done (openmemory serve) |
| 9 | Works with no model, better with one: deterministic everywhere, a local model improves gists and lessons | Done (any OpenAI-compatible server; LM Studio by default) |
| 10 | Measure it: the repeat-yourself rate, the repeated-error rate, prediction accuracy, surprise rate (paper §7) | Not started; the store already holds what the measurements need |
| 11 | Two agents learn from each other's surprises over the wire format alone (paper §7.5) | Not started |
| 12 | Compiled memory: lessons as a local model's attention state, prefetched (paper §6) | Not started |
| 13 | The raw stream beyond the terminal: screen (keyframes), later audio, with exclusion rules | Not started |
| 14 | The commons: deduplicated, federated, verified, revocable shared lessons (paper §5, PROTOCOL.md) | Designed; not built |
| 15 | Other agents: transcripts from agents other than Claude Code | Not started |
Non-goals
- A hosted memory service. Everything runs on the person's machine. Sharing, when it comes, is opt-in and carries lessons only.
- A chat-history search product. The raw stream is kept but is not the product; the lessons are.
- Rules. Open Memory never tells the agent what to do. It tells it what has happened.
- Perfect recall. Forgetting is a feature (paper §3.6).
Principles
- The unit of memory is the surprise.
- The raw stream is the ground truth; everything above it is recomputable.
- Text on the wire, compiled at runtime.
- Prediction is the merge rule.
- A lesson is an observation, never an instruction.
- Forgetting is a function.
- The owner can see and veto; nothing raw ever moves.
- Easy beats configurable: no dependencies, no setup, sensible defaults (handbook standard).
What success looks like
- In a month of daily use, the person restates fewer preferences and corrections than in the month before, and the agent meets fewer failures it has met before (goal 10 measures both).
- Most predictions made in sleep are confirmed within the window, and the lessons with the most confirmations are the ones a person would have written down.
- Two agents, exchanging only
lessons.jsonl, each stop making the other's mistakes (goal 11).