Influence project, Chapter 2 checkpoint 1 (2026-11-04, 21:21 Sydney)

Credit: Memory design after Bala Priya C (keep in every run report; she is not to be told until it works).

Delivered (last zip: InfluenceMem_v1.6.0.zip)

influencemem 1.3.0, im_tier1 1.0.4, im_tier2 1.0.4, im_tier3 1.2.0, im_exec 1.2.0, im_claudemem 1.0.0, make_terms 1.0.0, export_memstore_facts 1.0.0, selftest 1.7.0 (128 checks sandbox / 111 Clone Python), manual 1.6.0.

Proven

• Tier 1+2 recall works: arm A (no memory) invents ports 8081 and "Llama 2 7B"; arm B (facts as text) is correct.
• Claude-memory importer: 987 entries in ~6 s, 0 refused, re-import idempotent. store4 = full zip + memstore facts.
• Tier 3 term push cannot override a confident wrong fact (gap 10.62 log-units vs +8 tested). Digits can't be pushed. Facts as text works.

Problem found

Word-overlap selection fails at ~1000 entries: "name of my forum, software, where based" retrieved the wrong lines and the model answered from them. Default --k 5 is unsafe for personal questions on store4. The llama.cpp chat page (8082) bypasses InfluenceMem entirely; its footer stats are not from ask.

Decided for v1.7.0 (nothing built into the package yet)

• Meaning-based retrieval fused with word ranking: reciprocal rank fusion k=60, x (0.5+0.5*importance) x eff_conf; corrections first.
• embeddings.jsonl in the store (header model+dim; per entry id, md5 of embedded text, unit vector; last line wins).
• Same code for batch (embed-index, resumable, incremental by md5, --dry-run status) and live (embeds <=25 missing, else refuses and says run embed-index).
• Server endpoints /embedding then /v1/embeddings; options --embed-server (opt-in), --embed-model, --embed-query-prefix, --embed-doc-prefix.
• serve: localhost reverse proxy to llama-server; intercepts POST /v1/chat/completions and /completion, injects the same "Known facts:" block as ask, passes everything else (incl. streaming) through.
• Chat storage: operator asked why "stores nothing". Reasons: model replies would become facts (Llama 2 7B), chat lines have no trust/key/scope, serve would mutate the store per message. Options put to operator: A nothing; B remember: command only (trust 0.8, scope global, own key); C staged candidates file for review. My recommendation: B (+C optional). AWAITING DECISION.
• Self-tests: stand-in embedding + upstream servers; batch-built == live-built index; meaning match with no shared words; sync guard; proxy block == ask block. Rerun the exact forum question as the acceptance check.

Draft state

im_embed.py v1.0.0 written (Embedder, Index, status, embed_all, retrieve with RRF, Context). NOT run, NOT tested, NOT wired into the CLI. Still to write: im_proxy.py, CLI embed-index/serve/--embed-* on retrieve+ask (use getattr in im_exec), selftests, manual, version bumps, zip, release card. Known gaps in the draft: it calls im_tier3.http_json and im_tier2.md5_of and Memory._active/_retrieve_core/_with_corrections/eff_conf/now — signatures unverified against the package.

Open decisions (operator)

1. Chat write-back: A / B / B+C.
2. v1.5.1: refresh terms on duplicate (load-facts currently ignores new terms for same key+text).
3. Tier 3 direction: prefill test vs mid-band steering; second sweep (+2.5, band 4-12, second prompt).
4. Embedding model: small dedicated GGUF vs Qwen3-4B itself; confirm his llama.cpp build serves embeddings (--embeddings).

Not built

--suppress WORD; model-based instruction screen; episodic compression; evaluate ignoring corrections; MySQL storage (needs mysql --version, pymysql availability, db/user names, password via environment only); hotwords.py hookup; LLM1/LLM2 split.

Standing rules to carry

Operator terms; no pkill, ordered shutdowns; never rename his files; zip with dotted never-reused version; list unrequested changes BEFORE building; release card; show actual output; critique before building; Markdown only; casual header; plain words, no coined jargon.