bug: encoding agent rewrites memory content instead of faithfully encoding it

[CalebisGross]

Problem

When a memory is stored via remember, the encoding agent rewrites the content to something semantically unrelated. The stored raw content is correct, but the encoded summary/content diverges significantly.

Reproduction

Stored this content via remember:

"V7 dataset pipeline in progress (2026-04-09). Phase 1 complete: 1,182 raw inputs generated across 5 categories..."

The encoding produced this summary:

"Major architecture pivot: removed all generative LLL calls from mnemonic's cogni"

These are completely different topics. The encoding agent appears to be cross-contaminating with content from other memories or hallucinating.

Observed via

check_memory(raw_id="2326fc81-d7f9-4d6e-98b2-2578937bfc9c")
→ encoded memory_id: a221ca55-86c0-494c-a920-f66feb0d26bb
→ Summary: "Major architecture pivot: removed all generative LLL calls..."

Impact

This is a faithfulness failure in the encoding agent — the exact problem EXP-25 was designed to detect and fix. The encoding agent is producing summaries/content that don't reflect the actual input, which means memories retrieved via recall may not match what was actually stored.

Notes

This is likely caused by the current production spokes (EXP-20a checkpoint) which were shown to have faithfulness issues on diverse inputs (issue #381). The v7 dataset and retrained spokes should fix this once deployed.

[CalebisGross]

Fixed in #385. Root cause: getRelatedContext() used FTS5 keyword matching to find "related" memories and injected their summaries into the encoding LLM prompt. Lexical keyword overlap (e.g., "pipeline") pulled in semantically unrelated memories, and the 2B model latched onto the injected context instead of the actual input content.

Fix: removed getRelatedContext() entirely from the encoding pipeline. Episode context (temporal, same-session) is preserved — it provides genuine context. The related context feature was accidental complexity: encoding's job is faithful compression, not association-building. The advisory board analysis (Tesla, Musk, Hickey, Shannon, Faggin) unanimously agreed it was a cross-contamination vector with a self-reinforcing feedback loop (bad encodings contaminate future encodings).

Python training scripts (training_constants.py, prepare_faithfulness_data.py) updated to match — related_ctx parameter removed from build_production_prompt() so training data matches production prompt format.

Existing bad encodings from before this fix are still in the DB but won't propagate further.

[CalebisGross]

Closed via #385 (merged into feat/exp25-faithfulness-probe).