[PIR][WP3] Map LiveMem/TARL three-tier memory onto ruvllm context modules + ruvector-agent-memory
Nobody has claimed this yet.
- Dominant language
- Rust
- Stars
- 4.5k
- Forks
- 603
- Avg merge
- 23h 32m
- Merged PRs (30d)
- 59
Description
Epic: #837 · Plan: docs/research/perpetual-intelligence-runtime/03-program-plan.md
Goal
Formalize the three-tier memory architecture by mapping existing modules onto LiveMem's design, rather than building new tiers from scratch:
| Tier | Existing home |
|---|---|
| 1 — working context | ruvllm working_memory |
| 2 — continuous latent state | episodic_memory / semantic_cache / agentic_memory |
| 3 — transactional memory | ruvector-agent-memory (ADR-252) + reasoning_bank + ruvector-temporal-coherence (ADR-211) |
The work is making the tiering explicit, documented, and testable, not inventing storage.
Evidence
LiveMem — arXiv:2608.02515, submitted 2026-08-03. A fixed-capacity memory state whose lifetime is independent of the active context, maintained via a parallel memory branch (Gated DeltaNet-2 backbone) alongside a bounded KV attention window. Reframes long-running inference as computational state preservation rather than context growth: the system can answer from memory state after the supporting evidence has left the context window.
Disambiguation: cite as "LiveMem (arXiv:2608.02515)" — unrelated to the consumer iOS app of the same name.
Acceptance criteria
- Each tier has a documented interface, with the promotion/demotion policy between tiers stated explicitly.
- A test proves state survives context eviction — the system answers correctly from tier 2 after the supporting evidence has been evicted from working context. This is LiveMem's core claim and the single most important criterion here.
- Tier 2 has a hard fixed capacity enforced structurally; a test demonstrates it does not grow with context length over a long-running session.
- The existing module mapping above is verified against the code rather than assumed — if
episodic_memory/semantic_cache/agentic_memorydo not actually compose into a single fixed-capacity tier, say so and scope the delta. - ADR-252 and ADR-211 are cited by filename, not number alone —
ruvector's ADR set has ~15+ duplicate numbers under remediation in WP0a. - No persistence path is introduced that bypasses the proof-gated writes WP4 builds on.
Dependencies
WP1 (#847). Blocks WP4 (#840), and transitively WP12.
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with docs/research/perpetual-intelligence-runtime/03-program-plan.md and inspect the existing ruvllm working_memory, episodic_memory, semantic_cache, agentic_memory, ruvector-agent-memory, reasoning_bank, and ruvector-temporal-coherence modules. Verify whether the proposed mapping forms a fixed-capacity tier and identify the ADR-252 and ADR-211 filenames. Done means documented interfaces and promotion/demotion rules, tests for survival after context eviction and fixed capacity, and no bypass of proof-gated writes.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- rust
- Domain
- ai
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100