refactor(memory/v2): unify skills into concept-page activation pool#29619
Merged
Conversation
Skills no longer live in a separate Qdrant collection with a parallel activation pipeline. They are indexed into `memory_v2_concept_pages` under the slug prefix `skills/<id>` and flow through the same `simBatch` / `computeOwnActivation` / `selectInjections` pipeline as concept pages, with the same decay carry-over, top-K, and everInjected deduplication. The render path branches by slug prefix to keep the prompt format intact: skill slugs surface under `### Skills You Can Use`. The inspector activation log collapses concepts and skills into a single ranked array — the comparison is now apples-to-apples on the same scoring axis. `simSkillBatch`, `computeSkillActivation`, `selectSkillInjections`, the `memory_v2_skills` collection, and `top_k_skills` are gone. `top_k` defaults to 25 to preserve the prior 20 + 5 prompt budget. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5cb2b544b1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
siddseethepalli
added a commit
that referenced
this pull request
May 5, 2026
* refactor(memory/v2): unify rerank pool by pre-rerank A_o and apply boost additively Cross-encoder rerank now selects its top-K from the unified pre-rerank-A_o pool instead of running per-channel on top-K-by-fused-sim. Boost is added to A_o weighted by c_user / c_assistant rather than folded into sim_u / sim_a, so a slug strong in both channels can no longer crowd out single-channel hits via a doubled per-channel boost. simBatch loses its useRerank/rerankBoost options and becomes pure dense+sparse fusion. computeOwnActivation owns the rerank step end-to-end: it computes pre-rerank A_o per candidate, picks top-K, runs the cross-encoder once per channel against that unified set, and adds c_user·α·r_norm_u + c_assistant·α·r_norm_a to the final A_o. The macOS inspector now renders rerank Δ_u / Δ_a as standalone additive A_o rows. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(memory/v2): unify rerank pool by pre-rerank A_o and apply boost additively Cross-encoder rerank now selects its top-K from the unified pre-rerank-A_o pool instead of running per-channel on top-K-by-fused-sim. Boost is added to A_o weighted by c_user / c_assistant rather than folded into sim_u / sim_a, so a slug strong in both channels can no longer crowd out single-channel hits via a doubled per-channel boost. simBatch loses its useRerank/rerankBoost options and becomes pure dense+sparse fusion. computeOwnActivation owns the rerank step end-to-end: it computes pre-rerank A_o per candidate, picks top-K, runs the cross-encoder once per channel against that unified set, and adds c_user·α·r_norm_u + c_assistant·α·r_norm_a to the final A_o. The macOS inspector now renders rerank Δ_u / Δ_a as standalone additive A_o rows. Also drops stale `skills` references from the concept-frequency test — unblock the type-check that breaks on main after #29619 unified skills into the concept pool. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Vellum Assistant <assistant@vellum.ai> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 task
siddseethepalli
added a commit
that referenced
this pull request
May 9, 2026
…ed on cache miss (#30106) Address codex P1 review feedback on #29619: 1. Tag skill-seeded Qdrant points with payload.kind="skill" and have pruneSlugsWithPrefixExcept filter by that kind. validateSlug permits user-authored concept pages slugged `skills/foo`, which previously collided with the skill namespace and were repeatedly deleted by every seed run. 2. Exclude skill slugs whose in-process cache entry is missing from everInjected. Otherwise a startup race (Qdrant row present before the skill cache is seeded) marks the slug injected even though renderInjectionBlock silently dropped it, and per-turn deltas never re-attempt the attach until compaction. Co-authored-by: Vellum Assistant <assistant@vellum.ai>
Contributor
Author
|
Addressed in #30106 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
memory_v2_concept_pagesQdrant collection underskills/<id>and flow throughsimBatch/computeOwnActivation/selectInjectionsexactly like concept pages — same decay, same spread, same top-K, same everInjected dedup.simSkillBatch,computeSkillActivation,selectSkillInjections, thememory_v2_skillscollection, and thetop_k_skillsconfig are removed;top_kdefaults to 25 to preserve the prior 20 concepts + 5 skills budget.conceptsarray — skill rows appear with theskills/slug prefix, so the inspector compares the same scoring axis end-to-end.Original prompt
it