neuron-technologies/neuron
3
0
Fork 0
Code Issues 44 Pull Requests 7 Actions Packages Projects Releases Wiki Activity

feat(recall): activation-seed improvements [issues 2-3,5-10] #53

Closed
will.anderson wants to merge 0 commits from improve/recall-activation-seed into main
pull from: improve/recall-activation-seed
merge into: :main
Conversation 0 Commits - Files Changed -
will.anderson commented 2026-06-22 18:18:27 +00:00
Owner
Copy Link
Copy Source

Summary

Fixes activation-seed quality in handle_chat, handle_chat_agentic, handle_dharma_room_turn, and the nlg soul path.

Issues fixed

  • Issue 2: Brittle 50-char continuation threshold replaced with engram_is_continuation() (semantic detection using pronouns, continuation openers, length)
  • Issue 3: handle_chat and handle_chat_agentic used hist_len-1 (always the last assistant entry) as the topic anchor — scan-back now finds the prior user turn
  • Issue 5: Agentic seed now appends workspace_root so activation is workspace-aware
  • Issue 6: handle_dharma_room_turn and handle_dharma_room_turn_agentic passed the full transcript to engram_compile(); distill_transcript() now extracts the salient tail + question instead
  • Issue 7: dist/soul-with-nlg.el called engram_compile(message) with zero thread enrichment; now uses build_activation_seed() for both handle_chat and agentic paths
  • Issue 8: Semantic continuation detection via engram_is_continuation() in both paths
  • Issue 9: Old code sliced from position 0 (capturing preamble); topic_snip_from_entry() takes the TAIL of long assistant replies trimmed to last sentence boundary
  • Issue 10: Non-continuation messages now blend up to 3 prior user turns via multi_turn_topic() for richer thread continuity

New functions added to chat.el

  • topic_snip_from_entry(content) — tail-biased 200-char snip with sentence boundary
  • multi_turn_topic(hist, hist_len) — collect up to 3 prior user turns into seed
  • distill_transcript(transcript) — recency tail + question context from transcript
  • build_activation_seed(message, hist, hist_len) — central orchestrator for all fixes

Test plan

  • Short follow-up messages ("go on", "what else?") activate on prior user topic, not raw message
  • Continuation detection does not fire for short but topically new messages ("sleep", "AWS")
  • Agentic context includes workspace root in activation seed when set
  • Dharma room seed is focused (tail+question), not the full transcript
  • nlg soul path uses thread context instead of bare message
  • Multi-turn threads (3+ exchanges) recall concepts from earlier in the session
## Summary Fixes activation-seed quality in `handle_chat`, `handle_chat_agentic`, `handle_dharma_room_turn`, and the nlg soul path. ### Issues fixed - **Issue 2**: Brittle 50-char continuation threshold replaced with `engram_is_continuation()` (semantic detection using pronouns, continuation openers, length) - **Issue 3**: `handle_chat` and `handle_chat_agentic` used `hist_len-1` (always the last *assistant* entry) as the topic anchor — scan-back now finds the prior *user* turn - **Issue 5**: Agentic seed now appends `workspace_root` so activation is workspace-aware - **Issue 6**: `handle_dharma_room_turn` and `handle_dharma_room_turn_agentic` passed the full transcript to `engram_compile()`; `distill_transcript()` now extracts the salient tail + question instead - **Issue 7**: `dist/soul-with-nlg.el` called `engram_compile(message)` with zero thread enrichment; now uses `build_activation_seed()` for both handle_chat and agentic paths - **Issue 8**: Semantic continuation detection via `engram_is_continuation()` in both paths - **Issue 9**: Old code sliced from position 0 (capturing preamble); `topic_snip_from_entry()` takes the TAIL of long assistant replies trimmed to last sentence boundary - **Issue 10**: Non-continuation messages now blend up to 3 prior user turns via `multi_turn_topic()` for richer thread continuity ### New functions added to `chat.el` - `topic_snip_from_entry(content)` — tail-biased 200-char snip with sentence boundary - `multi_turn_topic(hist, hist_len)` — collect up to 3 prior user turns into seed - `distill_transcript(transcript)` — recency tail + question context from transcript - `build_activation_seed(message, hist, hist_len)` — central orchestrator for all fixes ## Test plan - [ ] Short follow-up messages ("go on", "what else?") activate on prior user topic, not raw message - [ ] Continuation detection does not fire for short but topically new messages ("sleep", "AWS") - [ ] Agentic context includes workspace root in activation seed when set - [ ] Dharma room seed is focused (tail+question), not the full transcript - [ ] nlg soul path uses thread context instead of bare message - [ ] Multi-turn threads (3+ exchanges) recall concepts from earlier in the session
will.anderson added 3 commits 2026-06-22 18:18:34 +00:00
feat(recall): recall-completeness improvements a60b1967df 
- Multi-query decomposition: split on AND/also/plus for multi-topic messages
- Named entity extraction: dedicated per-entity searches for project names
- Recall intent detection: boosted search pool for explicit recall requests
- Expanded pools: activation depth 8 (was 5), search 30->12 ranked (was 20->8)
- Threshold 25->15: retain moderately-relevant older nodes
- Sentinel cleanup extended to c14 for larger node pools
- Safe JSON truncation: find last closing brace before budget cap (8000 chars)
- Semantic continuation: engram_is_continuation replaces brittle 50-char threshold
- Thread snip: 150->250 chars for better pronoun resolution context
- Session preload: add project-specific and session-summary searches
feat(recall): context-dedup improvements dfa2a33926 
- Cache bell node result in engram_compile state (engram_compile_bell_node)
  so handle_chat affective_prefix reads the cached value instead of firing
  a duplicate engram query for distress signals (Issue 2)

- Cache primary activation result in engram_compile state
  (engram_compile_activation_json) using nodes0 from engram_compile_multi

- Replace redundant engram_activate_json(message, 2) in strengthen_chat_nodes
  with state_get(engram_compile_activation_json) - eliminates a third
  activation query per turn (Issue 7)

- engram_compile already has object-boundary truncation and cross-set
  dedup via engram_nodes_merge/engram_dedup_nodes (Issues 1, 6, 9)
feat(recall): thread-aware activation seed for nlg soul path [issue 7]
Neuron Soul CI / build (pull_request) Successful in 4m37s
Details
be02fcd960
will.anderson closed this pull request 2026-06-22 18:31:27 +00:00

Pull request closed

Please reopen this pull request to perform a merge.
Sign in to join this conversation.
Reviewers
No Reviewers
Labels
Clear labels
BETA-CRITICAL
On the beta critical path — must resolve before ship
 blocks-public-beta
Must be fixed before any public beta
 HELD
Blocked pending an external decision (counsel / dual sign-off)
 ORTHOGONAL
Parallel workstream (testing/hardening) — safe to assign off critical path
 P0
Drop-everything severity
 POST-BETA
Important but not blocking the beta ship
 security
Security vulnerability
No labels
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
andre.almeida (Andre Almeida) tim.lingo (Tim Lingo) will.anderson (Will Anderson)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: neuron-technologies/neuron#53
Delete Branch "improve/recall-activation-seed"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?