One memory across your AI tools — and every session opens already briefed: current goal, next action, what changed while you were gone. Local. Auditable. We couldn't read it if we tried.
This is the product. The capsule, the receipt, and the conflict alert are real hook output.
Using an AI agent? forget's llms.txt has setup instructions for LLMs.
LongMemEval, the standard long-term-memory benchmark. Full 500-question set, GPT-4o reader and judge for every row.
Knowledge-update questions — where memory products usually fail because stale facts survive: 92.3% (cloud-observer config)
Baseline and ceiling are from the LongMemEval paper: full-context is GPT-4o reading the entire transcript with no memory system; oracle is GPT-4o handed only the evidence sessions. The local row builds and retrieves memory entirely on-device (Qwen 14B observer, v3 reader) — run three times with different seeds: 78.4, 78.8, 78.0. We publish the spread; nobody else in this market does. The optional cloud-observer config (GPT-4o) reaches 81.8%. Our weakest category is single-session-preference — it's in the runs too. Per-question outputs and configs →
Three surfaces the hooks put in front of you — real output, not concept art.
The capsule: your goal, next action, parallel work, and any claim that's gone stale.
About to act on a fact that was later corrected? Both versions, labeled, before you commit.
Config, rules, hooks, scope, and a live MCP round-trip — checked end to end.
One local server, one connect command. Your existing MCP setup is
preserved and backed up; disconnect reverses everything.
Decisions surface from your conversations and your git history. The observation gate keeps facts and forgets junk — nobody writes documentation.
A context capsule opens every session with your open work. Relevant memories are pushed mid-conversation — and a conflict alert shows both versions when a fact was later corrected.
Claude Code, Codex, or Claude Desktop — the answer comes back with the commit or conversation that decided it. Sync arrives sealed E2EE · IN DESIGN
The three-minute proof, once installed: start a task, kill the session, reopen. A stateless agent greets you — yours opens with a handover: goal, next step, what changed. That moment is the product.
A coding agent (GPT-5.6, Codex) used forget across a multi-day trading-system build — data pipelines, safety constraints, live processes — and filed a structured review: acceptance criteria, regression scenarios, scores.
“Resuming long-running work: 8/10. The most valuable layer isn't generic recall — it's task state: progress, next actions, blockers, and the safety boundaries we agreed on, restored across sessions.”
Every defect it reported was reproduced, filed, and fixed — two releases within 24 hours. The issues are public →
Your model provider already reads your prompts. There is no reason your memory layer should be a second reader.
A compounding one — and today it lives on other people's servers.
In February 2026, a US federal court ruled that a defendant's AI conversations were not protected — the provider's terms said data could be used for training and disclosed. What you tell an AI is, legally, what you told a company.
U.S. v. Heppner, S.D.N.Y. 2026Rewind promised local-first privacy, pivoted to the cloud, and was acquired — its users were migrated onto the acquirer's terms overnight. Every startup holding your memories is one term sheet away from being someone else's database.
Rewind → Limitless → Meta, 2025A model provider cannot end-to-end encrypt memory its own servers must read, and it will never port your memory to a competitor's tool. Native memory features are silos by design.
structural, not a roadmap gap| forget | Mem0 | Zep | Native memory | |
|---|---|---|---|---|
| LongMemEval, full 500 | 78.4% ± 0.4 fully local · 81.8% best | self-reported, different readers — see methodology | unpublished | |
| Where your memory lives | your machine, always | cloud (self-host option) | their cloud | provider servers |
| Trust label on every read | ✓ | — | — | — |
| Decision receipts from git | ✓ | — | — | — |
| Session hooks: capsule · push recall · conflict alerts | ✓ | — | — | — |
| Follows you across tools | Claude Code · Codex · Claude Desktop | via MCP | via SDK | one vendor's silo |
| Open source | Apache-2.0 | ✓ | core only | — |
Competitor properties as publicly documented, July 2026. Corrections welcome — open an issue.
You're on the list. We'll write when sync ships. ✓
No spam — launch news only.
~/.forget —
storage, search, and memory building all run on your machine, offline
included. Local isn't a demo mode; it's a configuration we benchmark:
the fully-local pipeline scores 78.4% ± 0.4 on LongMemEval — three seeds,
spread published (81.8% with a GPT-4o observer), against the paper's
60.6% full-context baseline.
There is no account, no telemetry, and no server of ours in the loop.npx forget-connect disconnect
restores your original configs from backups and removes the hooks — third-party
settings are preserved byte-for-byte. Then delete ~/.forget and
your memories are gone. No account to close, no export request, no retention
window. It's your file.Because memory that cannot be deleted is surveillance. Ours forgets on command.
Supersede, retire, delete — every fact in forget can be struck through, and the strikethrough keeps the history. The logo is not a joke about the product. It is the product.
Your users' memories are your liability. A memory layer you cryptographically cannot read can't leak and can't be compelled. We're taking design partners.
Talk to us