recall
Skill4 starsChanged 54 days agoMITView on GitHub
Prime on Eidet's read path before changing it — hybrid lexical+vector+abstraction fusion, UCB exploration, dual-clock FadeMem recency, per-repo alpha learning, graph-neighbour and cue-anchor expansion, trust/ROI/quarantine de-boosts, per-type budgets, the generation-token recall cache, and L0/L1 context assembly. Use when the task touches RecallScoring, Fuse, RecallCache, RecallOptions/MemoryQuery, GetContextAsync, the wake-up budget, stage or valence recall filters, Memories_Search, or "why did this memory (not) come back". Not for the write-time gates or the trust model itself (see writepath), not for the MemoryEntry shape (see memory), not for retrieval benchmarking (see quality).
153 other projects use a version of this.
Adopted by
Most recently taken up by
- TencentBlueKing/bk-litetheir own version, first seen Nov 16, 2026
- TechNickAI/hermes-skillstheir own version, first seen Nov 16, 2026
- GoBeromsu/craft-skillstheir own version, first seen Nov 16, 2026
- gfunc/skillstheir own version, first seen Nov 16, 2026
- DeepExtrema/Tesseracttheir own version, first seen Nov 16, 2026
- shiwenbin1617/pstacktheir own version, first seen Nov 16, 2026
Used with
Projects with it also use addskill, apply-firmware and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
ryanportfolioSkilladdskill
Create, import, update, or install repo or personal skills: runtime ownership, resources, discovery checks.
Also in 6 projects that use this
ryanportfolioSkillapply-firmware
Bring current folder (plain folder, existing repo, old or partial firmware copy) up to full Harness Firmware in place. Mirror external repo → adopt-repo.
Also in 6 projects that use this
- +10 more from ryanportfolio
Other versions153 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 153
- Yeachan-Heo/oh-my-claudecode★ 41ksame name
- Yeachan-Heo/oh-my-claude-sisyphus★ 41ksame name
- backnotprop/pstack★ 4.2ksame name
- parcadei/Continuous-Claude★ 3.9ksame name
- parcadei/Continuous-Claude-v2★ 3.9ksame name
- parcadei/Continuous-Claude-v3★ 3.9ksame name
- indranilbanerjee/digital-marketing-pro★ 934same name
- ArtemXTech/personal-os-skills★ 545same name
- entireio/skills★ 240same name
- termdock/termdock-issues★ 160same name
143 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository12 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- canonSkill
- enrichmentSkill
- intakeSkill
- looseendsSkill
- maintenanceSkill
- memorySkill
- memorytoolSkill
- portalSkill
- qualitySkill
- sharingSkill
- writepathSkill
The file
Files
Recall & context — priming
Canonical spec: docs/domains/recall.md — read it for the full pipeline order, all invariants, key
files, and gotchas. Terms of record: UBIQUITOUS_LANGUAGE.md § Retrieval & context loading. Design
rationale: docs/specs/CoreSpec.md.
Two ranked surfaces, deliberately not unified: Recall (hybrid query → fused → policy → budgeted → cached) and Context (L0 identity + L1 dense-packed one-liners, <600 tokens). writepath defines trust; this domain only multiplies by it.
Core invariants (get these right)
RecallScoring.Fuseis the only home of the ranking math, and production policy stays outside it — trust, ROI, layer de-boost, and quarantine are applied afterFusebecause the benchmark callsFusedirectly.- Every field that changes the result set belongs in
RecallCache.ComputeKey— including the rounded alpha bucket and both expansion flags. A missing filter serves another query's cached results; this has bitten three times. - Resolve alpha before
TryGet(it's part of the key), andSetdrops the write if any tracked scope generation moved during the query. That is the coherence guarantee. FunctionalStage.Noneis a wildcard — the filter isStage == requested OR Stage == None. Filter clauses need explicitAndAlso(); RavenDB's default would widen them.- Three arms, and α blends only two of them. α is lexical-vs-vector; the abstraction arm adds
β·normAbson top. An absent third arm scores identically to two-arm fusion — which is what keeps the benchmark stable and lets a store fake skip the arm entirely. - Both expansions trust the loaded entry, not the pointer: in-scope by real
RepoIdandValidUntil is null. AnIsLatest-only check resurfaces forgotten memories. Cue expansion re-checks the store's own results — a lax backend must not be able to leak through recall. - Each reachability path needs its own
IntegrityCheck, and every probe pins the flags of the paths it isn't testing. A new expansion without one fails the auditor's coverage guard. - L1 carries no Observations (
[Insight, Procedure, Heuristic]only) — a failure stored as an Observation never resurfaces at wake-up. - Procedures are hard-capped in the wake-up; freed slots backfill insights, never heuristics.
- Type budgets are clamped to what the pool holds. A share is not a reservation: fixed shares left 23 of 87 repos rendering 13 of 20 lines because they owned no procedures or heuristics. The remainder goes to insights only; both hard caps still bind.
- An oversized L1 candidate is
continue, neverbreak. An unenriched memory renders its whole body, and one 28,890-character line truncated a real repo's wake-up to 6 of 20 lines. - Consolidation is capped at 6 of 20 wake-up slots, matched on
Source. Its output is redundant by construction — a scheduled re-consolidation of one cluster restates the same claim in new words — and theL1DuplicateThresholdword check cannot see a paraphrase (97% of duplicate wake-up lines were consolidation output at median word overlap 0.25). The cap is asymmetric on purpose; do not generalise it to all sources. The L1 candidate pool is 120 so the freed slots can backfill. - De-boost, never hide — every policy term is multiplicative.
Key files / reuse
src/Eidet.Core/Memory/RecallScoring.cs—Fuse,ExpandNeighbors,ExpandEntities,ApplyTypeBudgets,ComputeL1Score.src/Eidet.Core/Indexes/Memories_Search.cs—SearchVector(whole entry) andAbstractionVector(derived one-liner,IsNullOrEmpty-chained so a redacted field falls through); same embedder.src/Eidet.Core/Memory/RecallCache.cs— the key and the generation tokens.src/Eidet.Core/Services/MemoryService.cs—RecallAsync,GetContextAsync, the policy layer.src/Eidet.Core/Memory/MemoryRoi.cs,src/Eidet.Core/Maintenance/FadeMemCurve.cs,src/Eidet.Core/Layers/LayerScope.cs— the multipliers.
Gotchas
ExplainRecallAsyncis pre-expansion, cache-bypassing, hook-free — neither graph neighbours nor cue matches ever show up.- Cue expansion is only as dense as enrichment:
Entitiesis LLM-extracted, so the path is a no-op on an unenriched corpus. The abstraction arm is not — it falls back toContentat index time. - Two recency curves on purpose: fixed 7-day for L1, per-type
FadeMemCurvefor fusion. Don't unify. - Valence glyphs exist only in
RecallToolHandler; the wake-up has no stance marker and no dead-end floor (the designed negative-valence floor was never shipped). - A denying
PreRecallhook returns an empty list, indistinguishable from "no matches". - Access-count writes on the read path are intentional, patch-only, and exempt from invalidation — cached recalls are slightly stale on recency by design.
- Type budgets are soft: a backfill pass tops results up to
Limit.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

