neuron-technologies/el
3
0
Fork 0
Code Issues 4 Pull Requests 1 Actions 4 Packages Projects Releases 5 Wiki Activity

fix: engram search latency — pin embed model, cache query embeddings, bound activate BFS #105

Closed
will.anderson wants to merge 1 commits from worktree-agent-a7e7a591a07291058 into dev
pull from: worktree-agent-a7e7a591a07291058
merge into: :dev
Conversation 0 Commits 1 Files Changed 1
+140 -48
will.anderson commented 2026-08-15 19:25:25 +00:00
Owner
Copy Link
Copy Source

Preserving uncommitted work found in an open worktree during a worktree audit. Pins the Ollama embed model resident (keep_alive:-1) to avoid multi-second cold reloads (measured up to ~2.2s cold vs ~0.02-0.05s warm). Adds a direct-mapped query-embedding cache (FNV-1a keyed) so repeated queries cost zero Ollama round-trips. Replaces engram_activate's unbounded FIFO frontier BFS with a beam-capped, level-synchronous BFS (default 128, tunable via ENGRAM_ACTIVATE_BEAM) to bound hub-node explosion at depth 2-3. Applied from the worktree's own '.latency-deliverables/latency-fix.promotable.diff' — deliberately excludes an inert engram_prune_telemetry build-enabler stub the author noted was only needed to link this checkout, not part of the fix.

Preserving uncommitted work found in an open worktree during a worktree audit. Pins the Ollama embed model resident (keep_alive:-1) to avoid multi-second cold reloads (measured up to ~2.2s cold vs ~0.02-0.05s warm). Adds a direct-mapped query-embedding cache (FNV-1a keyed) so repeated queries cost zero Ollama round-trips. Replaces engram_activate's unbounded FIFO frontier BFS with a beam-capped, level-synchronous BFS (default 128, tunable via ENGRAM_ACTIVATE_BEAM) to bound hub-node explosion at depth 2-3. Applied from the worktree's own '.latency-deliverables/latency-fix.promotable.diff' — deliberately excludes an inert engram_prune_telemetry build-enabler stub the author noted was only needed to link this checkout, not part of the fix.
will.anderson added 1 commit 2026-08-15 19:25:27 +00:00
Fix engram search latency: pin embed model, cache query embeddings, bound activate BFS
El SDK CI - dev / build-and-test (pull_request) Failing after 10m1s
Details
1dc49b1923 
Pins the Ollama embed model resident (keep_alive:-1) to avoid multi-second
cold reloads whenever a larger generation model evicts it under unified-
memory pressure (measured cold reload up to ~2.2s vs ~0.02-0.05s warm).

Adds a direct-mapped query-embedding cache (FNV-1a keyed, full strcmp to
reject collisions) so a repeated query costs zero Ollama round-trips —
directly serves the curiosity loop, which reseeds the same query terms
repeatedly.

Replaces engram_activate's unbounded FIFO frontier BFS with a beam-capped,
level-synchronous BFS (default beam 128, tunable via
ENGRAM_ACTIVATE_BEAM) to bound per-hop hub-node explosion that could
previously reach multi-second/crash territory at depth 2-3.

Excludes an inert engram_prune_telemetry build-enabler stub that was only
needed to link this checkout against a newer integration branch — not part
of the fix.
will.anderson closed this pull request 2026-08-15 22:07:38 +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
 HELD
Blocked pending an external decision (counsel / dual sign-off)
 ORTHOGONAL
Parallel workstream (testing/hardening) — safe to assign off critical path
 POST-BETA
Important but not blocking the beta ship
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/el#105
Delete Branch "worktree-agent-a7e7a591a07291058"

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?