[everest-an]

Add Awareness Local — local-first MCP memory for AI coding agents

Re-opening after the previous PR (#47) was closed: the linked repository 404'd for reviewers at the time. The public distribution is now everest-an/Awareness-Market, which is live and verifiable.

Field	Value
Category	memory
Repo	https://github.com/everest-an/Awareness-Market
License	MIT
Language	TypeScript / Node (MCP server)

What it is. A local-first MCP memory server for AI coding agents (Claude Code, Cursor, Cline, Copilot, OpenClaw, …). Memory is stored as plain Markdown and indexed with SQLite FTS5 plus local embeddings, fused with RRF (hybrid recall). Retrieval makes zero LLM calls and works fully offline.

Evidence. 96.0% Recall@5 on LongMemEval (480/500), measured through the real daemon retrieval pipeline — with a published, honest comparison (including where it loses, e.g. MemPalace's raw baseline beats it by 0.6 pts) and reproducible scripts: https://awareness.market/benchmarks

On the 50-star guideline (novel/unique exception). It is one of the few fully local, zero-API memory servers in this space, and its retrieval/memory layer is being proposed as a standard, ERC-8350 (Agent Memory State) — a category that currently has no other standardized entry here. Happy to withdraw if you'd prefer to wait until it clears the star bar.

Regenerated with node scripts/generate-readme.js.

[autogame-17]

Reviewed head 83b9527fd4ab5e494fc4d5614ac851a3eae55903, its complete diff and the current canonical upstream everest-an/Awareness-Market at 14f11d3172e6496146ad50a38c62c52657fd0be8.

The repository is now accessible, active, MIT-licensed, and neither archived nor disabled. This is not a repeat of the earlier unavailable-link finding. Please address the following before inclusion:

1. data/projects.json:1445 and generated README.md:85: the advertised 96.0% Recall@5 needs a working, pinned reproduction. The linked daemon-path benchmark runner imports ../../sdks/local/src/core/.... That directory is absent from the canonical tree; the actual implementation is under src/core/. The published entry point therefore cannot resolve its first internal import in a fresh checkout. In addition, the benchmark page names multilingual-e5-small, while the English dataset path selects the English model, which MODEL_MAP maps to Xenova/all-MiniLM-L6-v2.

   Please either remove the numerical claim from the curated description for now, or provide a working immutable benchmark revision, matching model/dataset/budget settings, and the corresponding per-question output supporting 480/500. Identify the result as self-reported on LongMemEval_S. The arXiv URL resolves correctly, but it is the LongMemEval benchmark paper, not an independent validation of Awareness's score. I have not rerun the 500-question benchmark and am not claiming the score has been disproved.

   The advertised reproduction entry point is:

   node benchmarks/longmemeval/run_f053_daemon_path.mjs

2. Refresh from current upstream main and resolve the integration conflict. GitHub reports this branch as CONFLICTING. Preserve current catalog entries, keep the addition in data/projects.json, and regenerate README using the current generator rather than maintaining the old generated format manually. The PR's own merge-base diff contains only README/data changes; older script versions on its base are not being attributed to this PR.

The below-50-star novelty exception remains a separate maintainer decision (12 stars at this check). ERC-8350 exists as Draft: Agent Memory State Registry; draft status should remain explicit if it is cited as the novelty rationale.

After the content and branch update, verify the committed result:

node scripts/generate-readme.js
node scripts/validate.js
git diff --check
node scripts/check-links.js

Actual successful workflow checks are also required before approval. No merge, closure, catalog edit, or workflow approval has been performed.

[autogame-17]

One additional, isolated schema issue found while rechecking head 83b9527fd4ab5e494fc4d5614ac851a3eae55903 on 2026-10-12:

The new Awareness Local record's tags array at data/projects.json:1448-1454 contains five tags; the current schema permits two or three. This is separate from the older base branch's inherited validation problems.

I copied only this new record onto clean current main 596c20ba06738d5d7fcb10afd9934e507b4b81bd, used its trusted generator, then ran its validator. The exact result was:

Validated 121 projects across 10 categories
Validation failed with 1 error(s):
- projects[120]: tags must contain 2-3 non-empty strings

When refreshing the branch, reduce this array to two or three descriptive tags (for example, local-first, mcp, agent-memory), then regenerate README and verify:

node scripts/generate-readme.js
node scripts/validate.js
git diff --check
node scripts/check-links.js

The prior review's numerical-claim and integration requests remain unresolved: the canonical upstream is still 14f11d3172e6496146ad50a38c62c52657fd0be8, where the advertised daemon benchmark imports the absent sdks/local/src tree. The paper URL and repository are accessible; neither is being reported as a dead link. I have not rerun the benchmark or claimed that the self-reported score is disproved. No merge or workflow approval has been performed.

[autogame-17]

Verified the refreshed catalog-only diff against current main. The entry now has three valid tags and omits the unsupported benchmark claim and benchmark-paper association. Generated README, validation, and link checks passed locally and in the two current GitHub Actions jobs.