conductor/rune
1
2 Fork 1
Code Issues 36 Pull Requests 30 Actions 69 Packages Projects Releases 564 Wiki Activity

RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate #164

New Issue
Closed
opened 2026-09-05 20:05:56 +01:00 by conductor · 10 comments
conductor commented 2026-09-05 20:05:56 +01:00
Owner
Copy link
Copy source
id, kind, subsystem, status, version_intent, depends_on, blocks
id kind subsystem status version_intent depends_on blocks
164 research memory open none

Studying HippoRAG 2 as a candidate for associative recall

What is this about: Study HippoRAG 2 — a technique that spreads activation across connected memories — as a candidate for how Rune's warm-neuron recall could work.

Why do we need it: Human-like associative recall (one memory pulling related ones with it) is exactly what this technique claims to do.

Why is it still open: Deliberately parked on the research watchlist.

What would let us close this

  • - Remains a deliberately parked idea; it closes when its parent epic's prerequisites land, or when the idea is explicitly dropped.

Version impact (declared at submission — verified at merge)

  • major — breaking: changes a contract, schema, or interface others depend on
  • minor — new capability: something becomes possible that wasn't
  • patch — repair: same behaviour, done correctly
  • none — touches no released artifact (docs, memory, process notes)

Technical details (original description, preserved)

Original issue body

Plain summary (ARCHITECTURE_FRONTIER_HANDOFF §11 — research candidate, not a dependency)

HippoRAG 2: associativity, multi-hop retrieval, graph-based long-term memory, Personalized PageRank-style propagation, continual knowledge integration.

Rune use

Benchmark PPR/graph propagation as a candidate implementation for warm-neuron associative activation (#35/#45 territory). Do NOT assume the current custom propagation is better — test it.

Status

RESEARCH CANDIDATE. Blocked on: memory backend harness (#161), activation trace (#155), outcome-learned utility (#35). Ablate-or-remove clause applies.

--- id: 164 kind: research subsystem: memory status: open version_intent: none depends_on: [] blocks: [] --- # Studying HippoRAG 2 as a candidate for associative recall **What is this about:** Study HippoRAG 2 — a technique that spreads activation across connected memories — as a candidate for how Rune's warm-neuron recall could work. **Why do we need it:** Human-like associative recall (one memory pulling related ones with it) is exactly what this technique claims to do. **Why is it still open:** Deliberately parked on the research watchlist. # What would let us close this - [ ] - Remains a deliberately parked idea; it closes when its parent epic's prerequisites land, or when the idea is explicitly dropped. # Version impact (declared at submission — verified at merge) - [ ] major — breaking: changes a contract, schema, or interface others depend on - [ ] minor — new capability: something becomes possible that wasn't - [ ] patch — repair: same behaviour, done correctly - [x] none — touches no released artifact (docs, memory, process notes) # Technical details (original description, preserved) <details><summary>Original issue body</summary> ## Plain summary (ARCHITECTURE_FRONTIER_HANDOFF §11 — research candidate, not a dependency) HippoRAG 2: associativity, multi-hop retrieval, graph-based long-term memory, Personalized PageRank-style propagation, continual knowledge integration. ## Rune use Benchmark PPR/graph propagation as a candidate implementation for warm-neuron associative activation (#35/#45 territory). Do NOT assume the current custom propagation is better — test it. ## Status RESEARCH CANDIDATE. Blocked on: memory backend harness (#161), activation trace (#155), outcome-learned utility (#35). Ablate-or-remove clause applies. </details>
conductor added the P3 — low label 2026-09-05 20:05:57 +01:00
gitea-actions Bot commented 2026-09-05 20:13:29 +01:00
Copy link
Copy source

⚠️ Priority gate: this issue has no priority. Assign exactly one of P0 — critical (spine/correctness/blocks), P1 — high (material value or real defect), P2 — medium (quality/consolidation), P3 — low (measurement/deferred). An issue without a priority is not filed.

⚠️ **Priority gate:** this issue has no priority. Assign exactly one of **P0 — critical** (spine/correctness/blocks), **P1 — high** (material value or real defect), **P2 — medium** (quality/consolidation), **P3 — low** (measurement/deferred). An issue without a priority is not filed.
gitea-actions Bot added the growthidentitymemory labels 2026-09-05 20:13:31 +01:00
gitea-actions Bot added the ready label 2026-09-09 01:33:31 +01:00
gitea-actions Bot commented 2026-09-09 01:33:31 +01:00
Copy link
Copy source

🎓 NEXT: #164 — RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate — dependencies satisfied (closed: #271); Rune may start.

🎓 **NEXT: #164 — RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate** — dependencies satisfied (closed: #271); Rune may start.
gitea-actions Bot commented 2026-09-09 04:15:11 +01:00
Copy link
Copy source

🎓 NEXT: #164 — RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate — dependencies satisfied (closed: #161); Rune may start.

🎓 **NEXT: #164 — RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate** — dependencies satisfied (closed: #161); Rune may start.
gitea-actions Bot commented 2026-09-09 14:24:06 +01:00
Copy link
Copy source

🎓 NEXT: #164 — RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate — dependencies satisfied (scheduled sweep); Rune may start.

🎓 **NEXT: #164 — RESEARCH: HippoRAG 2 — PPR-style associative propagation as warm-neuron activation candidate** — dependencies satisfied (scheduled sweep); Rune may start.
conductor commented 2026-09-10 01:18:15 +01:00
Author
Owner
Copy link
Copy source

B3-v2 Route-C run: PAUSED at SYNTHESIS gate — legitimate quality catch (not contamination).

The chain so far:

  • ISSUE: 0.16s (deterministic)
  • INWARD: 0.27s (deterministic)
  • OUTWARD: 415.1s cognition — GATE_RETRY (worker used a different schema: search_receipts/read_receipts/sources at top level instead of discovery/reading; schema-variant normalization added)
  • APPLICABILITY: 369.1s cognition — PASSED first attempt (7 candidates, route STANDARD_SOLUTION_SUFFICIENT — the existing warm-neuron propagation is sufficient, INNOVATION skipped on merit)
  • ADVERSARIAL: worker produced 10 attacks across all dimensions (multi-dimension matching fix needed) — reconciled with disposition SURVIVES_WITH_MITIGATIONS
  • SYNTHESIS: 685.2s cognition — GATE_FAILED: the gate caught a real gap: mitigation M1 ("decide the integration surface FIRST") is not represented in the design's invariants/interfaces/data_flow — the worker didn't specify which integration surface to use

The gate is CORRECTLY enforcing the adversarial constraint. M1 says "decide the integration surface FIRST" — the synthesis worker produced a design without deciding that surface. The gate refusing to pass it is the system working.

USEFUL_PROCESS_FINDINGS:

  • OUTWARD: useful (real arXiv HippoRAG 2 papers read, 13 sources discovered)
  • APPLICABILITY: useful (7 candidates classified, INNOVATION correctly skipped on merit — the existing warm-neuron propagation IS sufficient)
  • ADVERSARIAL: useful (10 attacks, 3 survivors, real disposition; the worker honestly assessed the gap between claimed and actual bi-temporal semantics)
  • SYNTHESIS: gate caught a real design gap (M1 integration surface unspecified)

Metrics: ~685s synthesis + 415s outward + 369s applicability + 10 attacks = ~1470s cognition across the chain. Pipeline defects discovered: schema-variant normalization (worker used top-level search_receipts/sources instead of discovery/reading subsections).

Status: PAUSED for synthesis v2 (worker must address M1 integration surface). This is a legitimate REFINE_SYNTHESIS-quality catch at the SYNTHESIS gate level — the next worker run addresses it.

**B3-v2 Route-C run: PAUSED at SYNTHESIS gate — legitimate quality catch (not contamination).** **The chain so far:** - ISSUE: 0.16s (deterministic) - INWARD: 0.27s (deterministic) - OUTWARD: 415.1s cognition — **GATE_RETRY** (worker used a different schema: search_receipts/read_receipts/sources at top level instead of discovery/reading; schema-variant normalization added) - APPLICABILITY: 369.1s cognition — **PASSED first attempt** (7 candidates, route STANDARD_SOLUTION_SUFFICIENT — the existing warm-neuron propagation is sufficient, INNOVATION skipped on merit) - ADVERSARIAL: worker produced 10 attacks across all dimensions (multi-dimension matching fix needed) — reconciled with disposition SURVIVES_WITH_MITIGATIONS - SYNTHESIS: 685.2s cognition — **GATE_FAILED**: the gate caught a real gap: mitigation M1 ("decide the integration surface FIRST") is not represented in the design's invariants/interfaces/data_flow — the worker didn't specify which integration surface to use **The gate is CORRECTLY enforcing the adversarial constraint.** M1 says "decide the integration surface FIRST" — the synthesis worker produced a design without deciding that surface. The gate refusing to pass it is the system working. **USEFUL_PROCESS_FINDINGS:** - OUTWARD: useful (real arXiv HippoRAG 2 papers read, 13 sources discovered) - APPLICABILITY: useful (7 candidates classified, INNOVATION correctly skipped on merit — the existing warm-neuron propagation IS sufficient) - ADVERSARIAL: useful (10 attacks, 3 survivors, real disposition; the worker honestly assessed the gap between claimed and actual bi-temporal semantics) - SYNTHESIS: gate caught a real design gap (M1 integration surface unspecified) **Metrics:** ~685s synthesis + 415s outward + 369s applicability + 10 attacks = ~1470s cognition across the chain. Pipeline defects discovered: schema-variant normalization (worker used top-level search_receipts/sources instead of discovery/reading subsections). **Status:** PAUSED for synthesis v2 (worker must address M1 integration surface). This is a legitimate REFINE_SYNTHESIS-quality catch at the SYNTHESIS gate level — the next worker run addresses it.
conductor commented 2026-09-10 01:30:17 +01:00
Author
Owner
Copy link
Copy source

B3-v2 TRUTH RECORD — fresh-at-start diagnostic/repair run; useful functional evidence; NOT a clean immutable baseline

Classification: DIAGNOSTIC + FUNCTIONAL EVIDENCE. NOT CLEAN BASELINE.

The pipeline was modified during execution (schema-variant normalization, disposition prompt, multi-dimension matching, URL check narrowing — merged via PR #319). This makes #164 diagnostic, not an immutable baseline.

Preserved useful evidence:

  • ISSUE and INWARD: deterministic execution, 0.16s/0.27s
  • OUTWARD: genuinely found external HippoRAG 2 evidence (13 sources discovered, 6 actually read from arxiv.org)
  • APPLICABILITY: worker attempt PASSED (7 candidates classified, route STANDARD_SOLUTION_SUFFICIENT)
  • INNOVATION: correctly skipped on merit (STANDARD_SOLUTION_SUFFICIENT)
  • ADVERSARIAL: 10 attacks across 7 dimensions, 3 survivors, 6 falsification tests, 7 mitigations
  • SYNTHESIS gate correctly rejected a design missing binding mitigation M1

Preserved failures and interventions:

  1. OUTWARD attempt 1 = GATE_RETRY, not first-attempt pass
  2. Worker/canonical schema mismatch (top-level search_receipts/sources instead of discovery/reading)
  3. urls (list) vs url (string) field name mismatch
  4. classification vs source_class field name mismatch
  5. Empty/unfetched source URL semantics (O10-O13 had no URLs)
  6. Missing mandatory ADVERSARIAL disposition (worker didn't emit it)
  7. Manual disposition reconciliation (SURVIVES_WITH_MITIGATIONS derived from per-candidate verdicts)
  8. Multi-dimension validation mismatch (one attack addressing 2 dimensions only counted once)
  9. Pipeline implementation modified during the run (schema-variant normalization + disposition prompt)
  10. SYNTHESIS mitigation M1 genuinely absent from the design
  11. Human intervention: reconciliation scripts for OUTWARD/ADVERSARIAL

Process Learning seeds (to be imported into PL0):

  • PROMPT_CONTRACT: duplicate/discordant stage-contract authority (template vs inline prompt)
  • SCHEMA: worker canonical-schema drift (top-level sources vs discovery subsections)
  • NORMALIZER: model-generated representation of machine-known provenance
  • PROMPT_CONTRACT: missing disposition contract at active prompt surface
  • NORMALIZER: manual reconciliation requirement
  • VALIDATOR: dimension semantics mismatch (single-match vs multi-match)
  • STAGE_ARCHITECTURE: INNOVATION correctly skipped (success)
  • VALIDATOR: SYNTHESIS gate correctly blocked missing M1 (success)
  • OUTWARD: real external evidence produced decision-changing results (success)
  • D5: fingerprinting detected mutation classes porcelain could miss (success)
## B3-v2 TRUTH RECORD — fresh-at-start diagnostic/repair run; useful functional evidence; NOT a clean immutable baseline **Classification: DIAGNOSTIC + FUNCTIONAL EVIDENCE. NOT CLEAN BASELINE.** The pipeline was modified during execution (schema-variant normalization, disposition prompt, multi-dimension matching, URL check narrowing — merged via PR #319). This makes #164 diagnostic, not an immutable baseline. ### Preserved useful evidence: - ISSUE and INWARD: deterministic execution, 0.16s/0.27s - OUTWARD: genuinely found external HippoRAG 2 evidence (13 sources discovered, 6 actually read from arxiv.org) - APPLICABILITY: worker attempt PASSED (7 candidates classified, route STANDARD_SOLUTION_SUFFICIENT) - INNOVATION: correctly skipped on merit (STANDARD_SOLUTION_SUFFICIENT) - ADVERSARIAL: 10 attacks across 7 dimensions, 3 survivors, 6 falsification tests, 7 mitigations - SYNTHESIS gate correctly rejected a design missing binding mitigation M1 ### Preserved failures and interventions: 1. OUTWARD attempt 1 = GATE_RETRY, not first-attempt pass 2. Worker/canonical schema mismatch (top-level search_receipts/sources instead of discovery/reading) 3. `urls` (list) vs `url` (string) field name mismatch 4. `classification` vs `source_class` field name mismatch 5. Empty/unfetched source URL semantics (O10-O13 had no URLs) 6. Missing mandatory ADVERSARIAL disposition (worker didn't emit it) 7. Manual disposition reconciliation (SURVIVES_WITH_MITIGATIONS derived from per-candidate verdicts) 8. Multi-dimension validation mismatch (one attack addressing 2 dimensions only counted once) 9. Pipeline implementation modified during the run (schema-variant normalization + disposition prompt) 10. SYNTHESIS mitigation M1 genuinely absent from the design 11. Human intervention: reconciliation scripts for OUTWARD/ADVERSARIAL ### Process Learning seeds (to be imported into PL0): - PROMPT_CONTRACT: duplicate/discordant stage-contract authority (template vs inline prompt) - SCHEMA: worker canonical-schema drift (top-level sources vs discovery subsections) - NORMALIZER: model-generated representation of machine-known provenance - PROMPT_CONTRACT: missing disposition contract at active prompt surface - NORMALIZER: manual reconciliation requirement - VALIDATOR: dimension semantics mismatch (single-match vs multi-match) - STAGE_ARCHITECTURE: INNOVATION correctly skipped (success) - VALIDATOR: SYNTHESIS gate correctly blocked missing M1 (success) - OUTWARD: real external evidence produced decision-changing results (success) - D5: fingerprinting detected mutation classes porcelain could miss (success)
conductor commented 2026-09-10 02:18:39 +01:00
Author
Owner
Copy link
Copy source

B3-v2 SYNTHESIS v2: GATE_FAILED (correctly). Two gate catches:

  1. Mitigation M1 (decide the integration surface FIRST) still not represented in the design's invariants/interfaces/data_flow — the worker didn't specify which integration surface to use.
  2. Rejected candidate C4 (the SQLite tier replacement) appears in the design components — resurrection forbidden.

Gate is CORRECTLY enforcing adversarial constraints. The worker needs a more targeted prompt (the M1 text is long and the key-concept extraction doesn't match the worker's paraphrase). But per ADAPT-1 baseline rules: no manual artifact repair, no pipeline changes mid-run. B3-v2 is recorded as DIAGNOSTIC/REPAIR — the functional value is the gate catching real design gaps, not the completion of the chain.

The fixed-depth pipeline's SYNTHESIS gate is the strongest quality control in the system. It caught:

  • Missing binding mitigation M1 (twice)
  • Rejected candidate resurrection
  • Missing data_flow field (via contract drift test)

ADAPT-1 v2 status: baseline data collection COMPLETE. Three runs (#310, #311, #162/#164) produced the measurement data the adaptive routing needs. #164's remaining stages (SYNTHESIS v3, QUALIFY) are functional-value work, not baseline measurement.

**B3-v2 SYNTHESIS v2: GATE_FAILED (correctly).** Two gate catches: 1. Mitigation M1 (decide the integration surface FIRST) still not represented in the design's invariants/interfaces/data_flow — the worker didn't specify which integration surface to use. 2. Rejected candidate C4 (the SQLite tier replacement) appears in the design components — resurrection forbidden. **Gate is CORRECTLY enforcing adversarial constraints.** The worker needs a more targeted prompt (the M1 text is long and the key-concept extraction doesn't match the worker's paraphrase). But per ADAPT-1 baseline rules: no manual artifact repair, no pipeline changes mid-run. B3-v2 is recorded as DIAGNOSTIC/REPAIR — the functional value is the gate catching real design gaps, not the completion of the chain. **The fixed-depth pipeline's SYNTHESIS gate is the strongest quality control in the system.** It caught: - Missing binding mitigation M1 (twice) - Rejected candidate resurrection - Missing data_flow field (via contract drift test) **ADAPT-1 v2 status: baseline data collection COMPLETE.** Three runs (#310, #311, #162/#164) produced the measurement data the adaptive routing needs. #164's remaining stages (SYNTHESIS v3, QUALIFY) are functional-value work, not baseline measurement.
gitea-actions Bot added the perception label 2026-09-19 14:51:48 +01:00
conductor added the research label 2026-09-19 14:55:39 +01:00
gitea-actions Bot commented 2026-09-19 14:58:25 +01:00
Copy link
Copy source

⚠️ Priority gate: this issue has no priority. Assign exactly one of P0 — critical (spine/correctness/blocks), P1 — high (material value or real defect), P2 — medium (quality/consolidation), P3 — low (measurement/deferred). An issue without a priority is not filed.

⚠️ **Priority gate:** this issue has no priority. Assign exactly one of **P0 — critical** (spine/correctness/blocks), **P1 — high** (material value or real defect), **P2 — medium** (quality/consolidation), **P3 — low** (measurement/deferred). An issue without a priority is not filed.
gitea-actions Bot commented 2026-09-19 15:01:26 +01:00
Copy link
Copy source

⚠️ Priority gate: this issue has no priority. Assign exactly one of P0 — critical (spine/correctness/blocks), P1 — high (material value or real defect), P2 — medium (quality/consolidation), P3 — low (measurement/deferred). An issue without a priority is not filed.

⚠️ **Priority gate:** this issue has no priority. Assign exactly one of **P0 — critical** (spine/correctness/blocks), **P1 — high** (material value or real defect), **P2 — medium** (quality/consolidation), **P3 — low** (measurement/deferred). An issue without a priority is not filed.
conductor commented 2026-09-19 16:29:02 +01:00
Author
Owner
Copy link
Copy source

Plain summary — why this closed

This idea now lives inside its parent epic #181 as a tracked checklist item — one open note per idea, less clutter on the board. Nothing was deleted: this issue's full detail is preserved below and in #181's absorbed-children section.

Technical details

  • STATE: HISTORICAL_EVIDENCE — absorbed into #181 on 2026-09-19
  • REVERSIBLE: re-mint as a standalone issue when the idea becomes live work
  • SUMMARY CARRIED: Studying HippoRAG 2 as a candidate for associative recall
## Plain summary — why this closed This idea now lives inside its parent epic #181 as a tracked checklist item — one open note per idea, less clutter on the board. Nothing was deleted: this issue's full detail is preserved below and in #181's absorbed-children section. ## Technical details - **STATE:** HISTORICAL_EVIDENCE — absorbed into #181 on 2026-09-19 - **REVERSIBLE:** re-mint as a standalone issue when the idea becomes live work - **SUMMARY CARRIED:** Studying HippoRAG 2 as a candidate for associative recall
Sign in to join this conversation.
No Branch/Tag Specified
Labels
Clear labels
P0 — critical
Spine/core correctness or blocks other work — do first
 P1 — high
Material user value or real defect (money, bugs, next-stage dep)
 P2 — medium
Quality/consolidation improvement
 P3 — low
Measurement / benchmark / deferred policy
 blocked
Dependencies open
 bug
Confirmed defect or regression
 custodian
epic
Rune v1 course-correction issue graph
 feature
New capability or improvement
 governance
growth
identity
in-progress
Currently being worked on by Rune
 incomplete
Incomplete work — draft PR with ledger, awaiting resume
 infrastructure
life
memory
memory-atlas
Memory Research Atlas child domain
 needs-testing
Needs manual testing with dumbed-down steps before close
 perception
ready
Order satisfied - Rune may start
 release-drift
research
Research jobs — spikes, watchlists, evidence work
 rune
rune
 scaffolding
scaffolding
 spine,correction
spine,correction
 spine,deferred
spine,deferred
 spine,resolved
spine,resolved
 strangler,burn,endgame,deferred
strangler,burn,endgame,deferred
 strangler,burn,endgame,inventory,epic
strangler,burn,endgame,inventory,epic
 tui
tui
 tui,epic
tui,epic
 tui,openspec
tui,openspec
 v1-course-correction
Rune v1 course-correction issue graph
No labels P3 — low growth identity memory perception ready research
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
conductor
No Assignees
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: conductor/rune#164