Project Memory Synapse

Project Memory Synapse helps coding agents remember the corrections that matter: it grows source-backed implications, keeps trust human-governed, and refreshes the next session with proof.

Comment

Inspiration

Every team has a correction that must survive the next handoff: a customer constraint, an accessibility rule, a failed release condition, or a decision that changed the plan. Yet coding agents often start from the newest note or search result. They can have plenty of context and still repeat the mistake the team already worked through.

What it does

Project Memory Synapse gives those corrections a life after the conversation. A judge-created isolated project source or a public source enters the project's isolated store. PMS grows typed implications from exact source spans—requirements, risks, affected work, validation, release conditions, and next actions—while keeping each new relationship proposed until a person decides whether to accept, reject, or defer it.

That decision changes what the next agent session can use. Accepted relationships become readable Wiki knowledge and appear in a Fresh Session rebuilt after client state is cleared. The reviewer can then run a bounded check, validate its recorded work episode, and create a receipt that ties together the source, candidate, human decision, accepted relationship, action, outcome record, and changed session.

How we built it

The submission is an isolated JavaScript and Node.js implementation with a Cloudflare Workers-compatible adapter and a D1-compatible isolated-store adapter. It records source provenance, derives controlled source-span-grounded implications, issues a single-use review capability, keeps Wiki and Fresh Session projections accepted-edge-only, and validates a bounded action before producing a canonical receipt.

How Codex and GPT-5.6 were used

Core isolated-submission work was performed in Codex with GPT-5.6. GPT-5.6 also has one bounded product role: a strict-schema recommendation for one Tier-D relationship. The normal judge path uses a checked-in, packet-hash-bound replay marked live_api_call: false; it needs no API key or API credits. Only an explicit human review can change trusted state.

What we learned

The hard part of project memory is not storing more notes. It is carrying the right, source-backed decision into the next piece of work while making the path visible enough to inspect.

Boundaries

This project describes an isolated route. It does not claim production PMS access, a live GPT-5.6 API call, general NLU, real-world helped/harmed learning, or hosted acceptance.

Built With

Share this project:

Updates