[ATLAS][Memory] Port Hindsight-style multi-path recall (semantic + BM25 + graph + temporal + RRF) #1011

Open
opened 2026-04-22 14:25:20 +00:00 by Rockachopa · 2 comments
Owner

Parent

Timmy_Foundation/hermes-agent#985

Depends on

Timmy_Foundation/compounding-intelligence#230

Source to Mine

  • vectorize-io/hindsight

What

Port the most useful part of Hindsight's memory architecture into Hermes-native memory retrieval:

  • semantic retrieval
  • lexical/BM25 retrieval
  • graph-aware retrieval
  • temporal retrieval
  • fusion with RRF or equivalent
  • rerank stage where justified by measured gain

The goal is not to clone the whole Hindsight product. The goal is to steal the retrieval architecture that survives long-horizon memory workloads.

Tasks

  • Define a pluggable retrieval interface with separate retrieval lanes
  • Implement at least four retrieval lanes: semantic, lexical, graph, temporal
  • Add fusion logic (RRF or equivalent) with traceable scores
  • Add optional rerank stage behind a measurable guard
  • Emit retrieval traces so failures can be inspected after a run
  • Benchmark the new pipeline against the current Hermes-native retrieval path

Acceptance Criteria

  • Four independent retrieval lanes exist and can be enabled/disabled separately
  • Fusion output includes inspectable per-lane contributions
  • Retrieval traces are available for failed and successful recall cases
  • A benchmark or evaluation report shows whether the pipeline improves long-horizon recall on our prompt matrix
  • The implementation lands without introducing a mandatory cloud dependency
## Parent Timmy_Foundation/hermes-agent#985 ## Depends on Timmy_Foundation/compounding-intelligence#230 ## Source to Mine - `vectorize-io/hindsight` ## What Port the most useful part of Hindsight's memory architecture into Hermes-native memory retrieval: - semantic retrieval - lexical/BM25 retrieval - graph-aware retrieval - temporal retrieval - fusion with RRF or equivalent - rerank stage where justified by measured gain The goal is not to clone the whole Hindsight product. The goal is to steal the retrieval architecture that survives long-horizon memory workloads. ## Tasks - [ ] Define a pluggable retrieval interface with separate retrieval lanes - [ ] Implement at least four retrieval lanes: semantic, lexical, graph, temporal - [ ] Add fusion logic (RRF or equivalent) with traceable scores - [ ] Add optional rerank stage behind a measurable guard - [ ] Emit retrieval traces so failures can be inspected after a run - [ ] Benchmark the new pipeline against the current Hermes-native retrieval path ## Acceptance Criteria - [ ] Four independent retrieval lanes exist and can be enabled/disabled separately - [ ] Fusion output includes inspectable per-lane contributions - [ ] Retrieval traces are available for failed and successful recall cases - [ ] A benchmark or evaluation report shows whether the pipeline improves long-horizon recall on our prompt matrix - [ ] The implementation lands without introducing a mandatory cloud dependency
Author
Owner

Opened PR #1028: #1028

Ported multi-path holographic recall with lexical/semantic/graph/temporal lanes, RRF fusion, trace capture, and a prompt-matrix benchmark that improved top-1 recall from 1 to 3 on the new long-horizon fixture set.

Opened PR #1028: https://forge.alexanderwhitestone.com/Timmy_Foundation/hermes-agent/pulls/1028 Ported multi-path holographic recall with lexical/semantic/graph/temporal lanes, RRF fusion, trace capture, and a prompt-matrix benchmark that improved top-1 recall from 1 to 3 on the new long-horizon fixture set.
Author
Owner

BLOCKER: An open PR already references this issue.

PR #1028"feat: port multi-path holographic recall (#1011)" — is open and already implements the full work:

  • Four retrieval lanes: lexical, semantic, graph, temporal
  • RRF fusion with per-lane trace capture
  • Trace inspection via fact_store(action="trace")
  • Prompt-matrix benchmark report
  • No mandatory cloud dependency

PR is on branch fix/1011, authored by the issue opener, and is the intended implementation path.

Per STEP35 FREE BURN protocol: "STOP if any open PR already references #1011."

Resolution:

  • Review and merge PR #1028 (work is already done)
  • Close #1011 as duplicate/completed
  • Do NOT create a step35/ branch — the work exists and is in-progress via the canonical PR
**BLOCKER:** An open PR already references this issue. PR [#1028](https://forge.alexanderwhitestone.com/Timmy_Foundation/hermes-agent/pulls/1028) — *"feat: port multi-path holographic recall (#1011)"* — is **open** and already implements the full work: - Four retrieval lanes: lexical, semantic, graph, temporal - RRF fusion with per-lane trace capture - Trace inspection via `fact_store(action="trace")` - Prompt-matrix benchmark report - No mandatory cloud dependency PR is on branch `fix/1011`, authored by the issue opener, and is the intended implementation path. **Per STEP35 FREE BURN protocol:** *"STOP if any open PR already references #1011."* **Resolution:** - Review and merge PR #1028 (work is already done) - Close #1011 as duplicate/completed - Do NOT create a step35/ branch — the work exists and is in-progress via the canonical PR
Sign in to join this conversation.
No Label
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Timmy_Foundation/hermes-agent#1011