Influence — Checkpoint 4 (closes the Memory & Influence chapter)

Written 04/11/2026. Supersedes checkpoint 3 (stale). Latest delivered build: InfluenceMem_v1.4.0.zip.

Goal

One executive (ask) that calls three tiers, credits Bala Priya C, and has no defects:

• Tier 1 — Andrew's memstore.py approved facts (SQLite, read-only from above).
• Tier 2 — Bala Priya C's agent-memory design: working, episodic, semantic and procedural layers; write policy; trust and provenance; expiry and decay; retrieval at decision points; corrections kept as procedural memory.
• Tier 3 — "Influence": a bounded, fading logit push on local Qwen3-4B (llama-server, port 8082, on the Clone).

Status

Part	State	Evidence
Tier 1 (facts in)	Working	export_memstore_facts.py writes JSONL because Python 3.15 on the Clone has no sqlite3
Tier 2 (memory, retrieval)	Working	Recall test: model alone invented ports 8081 and "Llama 2 7B"; with facts as text it gave 8080/8082 and Qwen3-4B correctly
Relevance floor	Working	min_relevance 0.2 removed the 4 unrelated facts that matched on common words
Executive ask (arms A–D)	Working	Verified on the Clone, 72 checks, 0 failed (v1.1.6); later versions 97/113 in sandbox
Tier 3 push, mid-band	Works, tuning only	See sweep below
MySQL storage	Not built (final cut)	Waiting on his mysql --version, whether pip install pymysql works, database and user names (password via environment variable only)
Hybrid retrieval (keyword + vector), LLM1/LLM2 split, hotwords.py hookup	Not built	Design from his Experiment 001/002

Mid-band sweep (one prompt, temperature 0, 120 tokens, arm A vs arm D, band 4-30)

	A	+2	+3	+4
Distinct-word ratio	0.48	0.58	0.65	0.60
Repeated 4-word runs	23%	1%	0%	1%
Reply words found in prompt	30%	54%	47%	42%

Reading: +2 and +3 are fluent, on topic and do not loop; +4 turns vague; +6 (earlier, with facts) drifted off topic. "More detailed yet broader" is partly supported at +2/+3 only. One run per level on one prompt is indicative, not proof. Part of the gain may be breaking the baseline's loop.

Open items for the next chapter

1. Put the on-topic share into summary.md as a calculated row (done by hand so far).
2. Second sweep: +2.5 and band 4-12, with a second prompt.
3. Write a terms.json for the memstore entries and test term pushes (e.g. 8082, Qwen3-4B). Do not write into his memstore.
4. MySQL: seam is _file, _load, _save, _audit, init_store, snapshot/rollback. Cannot be tested in the sandbox.
5. Not built: --suppress WORD, model-based instruction screen, episodic compression, evaluate ignoring corrections.
6. Bala Priya C's decision-tree article is unseen. He is not telling Bala until it works.

Facts a new chat must know

• Clone is CentOS 6. python3 = /opt/python-3.15.0 (no sqlite3); /usr/bin/python is Python 2.6 (no argparse). Qwen starts by itself after reboot; health at http://localhost:8082/health.
• One llama-server slot: one job at a time. Greedy decoding cascades, so any small prompt change reroutes everything.
• Standard prompt md5 d78a0cc67c870050941cdfc719e0a365 (536 words), in ~/memstore and ~/memstore/LogitStaircase.
• Split multi-byte characters crash naive clients; v1.1.6 onward fetches the whole character.
• Delivery rules apply: bump version and date, manual history, version shown at run, dotted zip, list unrequested changes first, release card.