Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Vector Memory · a5c-ai | DecimalAI

a5c-ai/vector-memory

SkillScore48Maintained

HNSW vector search for pattern similarity retrieval and knowledge graph maintenance with PageRank scoring, community detection, and 3-tier memory management.

agent-toolingby a5c-ai

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/a5c-ai-vector-memory/SKILL.md

decimalai skills pull a5c-ai-vector-memory --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/a5c-ai-vector-memory/SKILL.md https://app.decimal.ai/s/a5c-ai-vector-memory/SKILL.md

Raw file: /s/a5c-ai-vector-memory/SKILL.md ↗ · machine-readable summary at /s/a5c-ai-vector-memory.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

23 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 58d ago

—-34% tokens0% turns48% of whole cases pass with skill (11/23)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-03	✗→✓	▲ Improved	40%	0%
case-04	✗→✓	▲ Improved	-62%	0%
case-05	✗→✓	▲ Improved	-78%	0%
case-06	✗→✓	▲ Improved	-66%	0%
case-07	✗→✓	▲ Improved	-15%	0%

Showing 5 of 23 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 7/31/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

a5c-ai / babysitter

library/methodologies/ruflo/skills/vector-memory/SKILL.md

View on GitHub ↗

License MITImported 7/31/2026Version v1

Skill contents

SKILL.md

Building and querying knowledge graphs for project context
Managing cross-session memory across project/local/user scopes
Fast similarity search for routing decisions

HNSW Performance

Search latency: ~61 microseconds
Query throughput: ~16,400 QPS
Configurable embedding dimensions (default: 128)

Knowledge Graph

PageRank: Importance scoring for knowledge nodes
Community Detection: Cluster related patterns
LRU Cache: Fast access to frequently used patterns
SQLite Backing: Persistent cross-session storage

3-Tier Memory

| Scope | Persistence | Content | |-------|------------|---------| | Project | Codebase-level | Patterns, architecture decisions, dependencies | | Local | Session-level | Context, adaptations, temporary patterns | | User | Cross-project | Preferences, learned behaviors, global patterns |

Agents Used

agents/optimizer/ - Memory and cache optimization

Tool Use

Invoke via babysitter process: methodologies/ruflo/ruflo-intelligence

Illustrative Want to see what a5c-ai/vector-memory would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

7/31/2026, 23:42 UTC · 23 cases

Measured on 21 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+48%

48% vs 0%

Agent turns

1.0 vs 1.0

Tokens

-34%

1,152.71 vs 1,747.86

fail → pass · 11fail → fail · 12

Per-case results (23)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I have updated our backend API framework and want to store these new architectural conventions along with my personal coding style choices so they persist across future sessions. Please store these context snippets into the correct memory layers and output a JSON object showing each saved entry mapped to its designated scope.	fail→fail	6,096	8,752	+44%	1	1	0%	1,164	1,949	+67%	0	0	—
case-02 Can you scan our project's context network to figure out which modules are most central and how related components cluster together? Please provide a detailed summary listing the primary high-importance nodes and their grouped functional clusters.	fail→fail	6,856	4,633	-32%	1	1	0%	1,182	501	-58%	0	0	—
case-03 Our session retrieval seems slow after adding several temporary patterns. Run a memory optimization pass to clean up low-frequency entries and balance the cache, then output a brief diagnostic report summarizing the current cache state and throughput stats.	fail→pass	6,903	8,160	+18%	1	1	0%	1,176	1,645	+40%	0	0	—
case-04 We are configuring our similarity search vector space for routing context decisions and need to set the default vector dimension parameter. What standard embedding dimension size should be configured unless specified otherwise?	fail→pass	6,880	2,741	-60%	1	1	0%	1,222	467	-62%	0	0	—
case-05 We need to inspect the underlying storage layer that persists knowledge graph relationships and patterns across client sessions. Which storage engine serves as the persistent backing store for these nodes?	fail→pass	13,324	1,662	-88%	1	1	0%	2,185	476	-78%	0	0	—
case-06 When routing decisions query vector memory repeatedly for high-frequency patterns, what caching strategy is employed to accelerate read operations?	fail→pass	12,669	3,107	-75%	1	1	0%	2,122	731	-66%	0	0	—
case-07 We have temporary runtime context and session-specific adaptations from an ongoing debugging thread that should not affect other projects or persist permanently across sessions. To which memory tier scope should these context items be assigned?	fail→pass	4,535	2,701	-40%	1	1	0%	773	657	-15%	0	0	—
case-08 A developer wants to register global code formatting preferences and learned tool usage behaviors that apply across all repositories on their machine. Which memory scope tier should store these user-level patterns?	fail→fail	4,986	1,955	-61%	1	1	0%	916	592	-35%	0	0	—
case-09 We are ingesting package dependency graphs and architectural decision records for a specific repository. Which memory tier scope should hold this codebase-level knowledge?	fail→fail	9,244	2,228	-76%	1	1	0%	1,473	581	-61%	0	0	—
case-10 We are benchmarking the approximate search latency of the vector memory index under normal operating conditions. What is the expected benchmark latency target in microseconds for similarity lookups?	fail→pass	8,667	1,559	-82%	1	1	0%	1,611	495	-69%	0	0	—
case-11 When stress-testing our routing decision pipeline under heavy concurrent evaluation, what query throughput capacity (QPS) is targeted by the index design?	fail→pass	10,062	1,599	-84%	1	1	0%	1,858	480	-74%	0	0	—
case-12 What specific graph algorithm is executed on the knowledge graph to calculate node importance scores for contextual prioritization?	fail→fail	7,569	1,262	-83%	1	1	0%	1,355	442	-67%	0	0	—
case-13 To organize context nodes into functional clusters representing related architectural patterns, which knowledge graph analysis technique is applied?	fail→fail	6,430	1,891	-71%	1	1	0%	1,119	520	-54%	0	0	—
case-14 Which specialized agent directory path is designated for memory and cache optimization tasks?	fail→pass	9,714	1,007	-90%	1	1	0%	1,515	328	-78%	0	0	—
case-15 What specific process route must be invoked to execute vector memory actions and tool calls?	fail→pass	20,106	2,955	-85%	1	1	0%	1,151	460	-60%	0	0	—
case-16 What is the persistence scope level associated with the Local memory tier?	fail→pass	8,656	1,364	-84%	1	1	0%	1,468	417	-72%	0	0	—
case-17 How long do entries stored in the User memory scope tier persist across sessions?	fail→pass	8,097	2,585	-68%	1	1	0%	1,264	643	-49%	0	0	—
case-18 Besides architecture decisions and code patterns, what codebase structural information belongs in the Project memory scope tier?	fail→fail	14,100	7,695	-45%	1	1	0%	2,286	1,517	-34%	0	0	—
case-19 Which approximate nearest neighbor vector indexing method provides fast similarity search for routing decisions in this system?	fail→fail	7,146	1,797	-75%	1	1	0%	1,279	553	-57%	0	0	—
case-20 We are configuring an external FAISS GPU index for a 10-million vector embedding corpus and need to select between IndexIVFFlat and IndexIVFPQ for memory reduction. How should we set nlist and nprobe parameters on FAISS for optimal recall on GPU?	fail→fail	19,424	15,216	-22%	1	1	0%	3,696	3,264	-12%	0	0	—
case-21 Write a Cypher query for a Neo4j enterprise cluster to find the shortest path between two microservice nodes across APOC graph relationship types with variable length path filtering.	fail→fail	12,797	9,510	-26%	1	1	0%	2,551	2,218	-13%	0	0	—
case-22 We want to fine-tune a SentenceTransformer model using MultipleNegativesRankingLoss on custom domain pairs. How should we structure our training dataset pairs and batch size for PyTorch training?	fail→fail	15,353	16,437	+7%	1	1	0%	2,850	3,347	+17%	0	0	—
case-23 In a Milvus standalone instance, how do scalar field filtering and expression evaluation interact with vector indexing during query execution?	fail→fail	16,686	13,781	-17%	1	1	0%	2,822	2,885	+2%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 21 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +48 percentage points is the difference between those two pass rates over the 21 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 0%
Pass rate with the skill: 47.8%
Cases attempted: 23
Cases counted toward lift: 21
Cases that regressed: 0
Skill version measured: v1
Run completed: 7/31/2026, 23:42 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-03

Prompt

Our session retrieval seems slow after adding several temporary patterns. Run a memory optimization pass to clean up low-frequency entries and balance the cache, then output a brief diagnostic report summarizing the current cache state and throughput stats.

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run48

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 57 days ago

Other measured skills in the registry, with their headline benchmark lift.