Loop 2 layer 5: session-aware weighting #8

New issue

Closed

opened 2026-05-23 20:55:37 +00:00 by coilysiren · 0 comments

coilysiren commented

2026-05-23 20:55:37 +00:00

Owner

Originally filed by @coilysiren on 2026-05-20T12:01:07Z - https://github.com/coilysiren/voice-flow-learning-loop/issues/9

Context

Layer 5 of the 8-layer build path for Loop 2 documented in README.md.

Goal

A phrase appearing 50 times in one session because Kai was iterating on a bug is task-specific noise. A phrase appearing 50 times across 50 sessions is durable signal worth a snippet. Weight cross-session repetition higher than intra-session.

Pull session boundaries from session-lattice if its session-grouping is exposed via MCP. Otherwise fake it via repo-recall's session metadata (each session has a session_id in the JSONL filename and repo-recall already groups by session).
Add a session-spread metric to each candidate's output: "seen 47 times across 12 sessions" vs "seen 47 times in 1 session".
Ranking incorporates session-spread: value = frequency × length × log(session_count + 1) or similar.

Done when

Candidate output includes session-spread alongside frequency and length.
Ranking incorporates session-spread.
A candidate that occurs only in one session ranks below an otherwise-equivalent candidate spread across sessions.

Depends on: layer 4. Unblocks: higher signal-to-noise for layer 6 onward.

_Originally filed by @coilysiren on 2026-05-20T12:01:07Z - [https://github.com/coilysiren/voice-flow-learning-loop/issues/9](https://github.com/coilysiren/voice-flow-learning-loop/issues/9)_ **Context** Layer 5 of the 8-layer build path for Loop 2 documented in [README.md](../blob/main/README.md). **Goal** A phrase appearing 50 times in *one* session because Kai was iterating on a bug is task-specific noise. A phrase appearing 50 times across 50 sessions is durable signal worth a snippet. Weight cross-session repetition higher than intra-session. - Pull session boundaries from session-lattice if its session-grouping is exposed via MCP. Otherwise fake it via repo-recall's session metadata (each session has a `session_id` in the JSONL filename and repo-recall already groups by session). - Add a session-spread metric to each candidate's output: "seen 47 times across 12 sessions" vs "seen 47 times in 1 session". - Ranking incorporates session-spread: `value = frequency × length × log(session_count + 1)` or similar. **Done when** - Candidate output includes session-spread alongside frequency and length. - Ranking incorporates session-spread. - A candidate that occurs only in one session ranks below an otherwise-equivalent candidate spread across sessions. **Depends on:** layer 4. **Unblocks:** higher signal-to-noise for layer 6 onward.

coilysiren added the

label

2026-05-31 01:55:11 +00:00

coilysiren

2026-06-17 08:24:09 +00:00