Skip to content

thread-research-memory.md

Latest commit

 

History

History
176 lines (130 loc) · 4.23 KB

thread-research-memory.md

File metadata and controls

176 lines (130 loc) · 4.23 KB
Raw

Tutorial: Research memory with a Cognitive Thread

Goal: use one Thread to hold a hypothesis, claims, evidence quotes, and retrieval traces — without turning Paper_Rec into a paper-writing tool.

Sample data in-repo: content/threads/mm-llm-alignment/.

0. Setup

cd Paper_Rec_Skill
pip install -e packages/wiki-bridge

Optional Wiki UI: see CONTRIBUTING.md.

1. Inspect the sample thread

python -m wiki_bridge.cli thread-show --wiki-root . --id mm-llm-alignment

You should see:

Field Example
Hypothesis Unified objective for multimodal preference alignment
Claim C1 preference data quality dominates algorithm choice…
Gap empirical ablation on preference data quality
Papers e.g. llm/2025/getting-started

Ledger: content/threads/mm-llm-alignment/events.jsonl
Evidence map: content/threads/mm-llm-alignment/evidences.jsonl

2. Create your own thread (optional)

From template (recommended)

python -m wiki_bridge.cli thread-template-list --wiki-root . --seed
python -m wiki_bridge.cli thread-template-import \
  --wiki-root . \
  --template multimodal-alignment \
  --id my-mm-align

Or in Wiki UI: open 研究主线 →「主线模板市场」→ 导入.

From scratch

python -m wiki_bridge.cli thread-create \
  --wiki-root . \
  --title "My research direction" \
  --hypothesis "…" \
  --keywords "a,b,c"

Edit content/threads/<id>/thread.json to add claims, evidence_gaps, seed_queries.

3. Retrieve with thread context

In an agent that loads skill/:

thread:mm-llm-alignment
multimodal preference data quality vs algorithm ablations

Or enable iterative refine:

thread:mm-llm-alignment iterative
…

Skill Modules 1.5 → 2a/2b → 2.5 inject seeds/gaps, multi-path search, optional one refine wave, then Thread relevance R.

Report sections: Retrieval Trace + Thread relevance (skill/output-template.md).

4. Persist papers + optional retrieval trace

python -m wiki_bridge.cli sync-report \
  --wiki-root . \
  --report path/to/report.json \
  --thread mm-llm-alignment \
  --query-id demo-2026-07

If report.json contains retrieval_trace: [...], each round is appended as kind: query_iter.

Manual trace:

python -m wiki_bridge.cli query-trace \
  --wiki-root . \
  --thread mm-llm-alignment \
  --round 0 \
  --path-id gap \
  --query "preference data quality multimodal ablation" \
  --raw-hits 30 \
  --kept 12

5. Bind Claim–Evidence (quote → claim)

CLI

python -m wiki_bridge.cli thread-evidence-add \
  --wiki-root . \
  --thread mm-llm-alignment \
  --claim-id C1 \
  --path llm/2025/getting-started \
  --quote "…" \
  --stance supports \
  --suggested

Accept when you agree:

python -m wiki_bridge.cli thread-evidence-gate \
  --wiki-root . \
  --thread mm-llm-alignment \
  --evidence-id E1 \
  --gate accepted

Wiki UI: open a paper page → select a paragraph →「挂到主线」→ pick thread + claim. Thread detail shows the evidence panel.

6. Watch / Delta (optional)

python -m wiki_bridge.cli thread-delta \
  --wiki-root . \
  --id mm-llm-alignment \
  --mode gap_focus \
  --print-md

Briefs land under content/threads/mm-llm-alignment/deltas/.

7. Link an experiment (when you have one)

python -m wiki_bridge.cli thread-link-exp \
  --wiki-root . \
  --id mm-llm-alignment \
  --exp-id demo-ocr-handwriting-v1

Multi-run curves: put metrics/curves.json + metrics/curves_<run>.json, open Exp detail in Wiki (run overlay / ?compare= / poll).

Mental model

flowchart LR
  H[Hypothesis] --> C[Claims]
  C --> E[Evidences]
  E --> P[Wiki papers]
  C --> G[Evidence gaps]
  G --> Q[query_iter / retrieval]
  Q --> P
  C --> X[Experiments]
Loading

Out of scope here: LaTeX manuscripts, citation audit, PDF finalization (use tools like Anaxa for that). Paper_Rec owns the cognitive ledger until the metric moves.

Next reading