Technical Summary: Eywa – Provenance-Grounded Long-Term Memory for AI Agents

1. Problem Definition

AI agents operating across multiple sessions require memory systems that can retrieve, audit, update, and erase information. Current memory architectures often collapse source evidence, extracted facts, retrieved context, and answer policies into a single, opaque prompt path. This "black box" approach makes failure diagnosis difficult: a wrong answer could stem from missing evidence, unsupported extraction, stale state, retrieval loss, or the answer model's behavior, yet end-to-end benchmarks typically report only a single aggregate score.

Furthermore, simply extending the context window of Large Language Models (LLMs) is insufficient. Long-context models suffer from high computational costs at scale and often fail to reliably retrieve evidence located in the middle of prompts or when tasks exceed effective context sizes. Existing systems also struggle to distinguish between immutable source data and derived beliefs, making it hard to verify claims, perform user erasure, or repair specific memory errors without reprocessing the entire history.

2. Methodology: The Eywa Architecture

Eywa introduces a provenance-grounded memory architecture built on the principle of "evidence before belief." It separates the memory lifecycle into distinct, auditable stages: immutable evidence capture, validated belief extraction, deterministic retrieval, and policy-separated answer synthesis.

2.1. Core Objects and Invariants

Eywa defines three distinct memory objects, maintaining the invariant that evidence precedes belief:

1. Evidence: Immutable source material (raw user turns, timestamps, roles, metadata). This is the authoritative substrate.
2. Signals: Typed detections over evidence (dates, entities, decisions, corrections) extracted via deterministic detectors without LLM calls.
3. Beliefs: LLM-derived canonical facts or observations. These are valid only to the extent they are supported by source evidence and remain linked to it via provenance metadata.

2.2. The Write Path: Evidence Before Belief

The write process operates in two tiers to ensure grounded extraction:

• Tier 0 (Immutable Capture): Every user turn is stored as an immutable evidence row. A deterministic signal detector extracts typed anchors (e.g., dates, URLs, monetary values) to serve as a validation substrate.
• Tier 1 (Validated Extraction): An LLM proposes candidate facts. A validation function checks these candidates against the Tier 0 evidence and signals. It verifies:
  • Support: Lexical and fuzzy overlap with the source text.
  • Hard Anchors: Exact matching of deterministic values (dates, IDs, amounts).
  • Subject & Act: Verification that the claimed subject exists and that negation/uncertainty is preserved.
  • Result: Only validated facts are written as canonical beliefs, retaining a link to their source evidence. Failed candidates are rejected or marked for repair.

2.3. The Read Path: Deterministic Multi-Route Retrieval

Eywa's read path is LLM-free (zero LLM calls occur between query receipt and context return). It uses a deterministic query planner to analyze the question's shape (e.g., inference, list, temporal) and assigns weights to multiple retrieval routes:

• Vector Search: Embedding cosine similarity.
• Keyword Search: SQLite FTS5 BM25 matching.
• Temporal Search: Filtering by fact validity windows.
• Entity/Graph Search: Traversing knowledge graphs based on extracted entities.
• Support Channels: Raw-episode rescue for exact questions and inference support for reasoning tasks.

Candidates from these routes are fused using Weighted Reciprocal Rank Fusion (RRF). Post-fusion filters apply person demotion (penalizing facts about the wrong entity) and a preservation floor (protecting high-ranked candidates from being buried by noise). The final ranked list is packed into a bounded context string under a token budget.

2.4. Answer Policy Separation

The system returns two distinct fields to the answer model:

1. Context: The retrieved memory evidence.
2. Answer Instructions: Explicit policy text (e.g., strict, balanced, advanced) dictating how the model should use the evidence.
   This separation allows the same memory substrate to be evaluated across different model regimes (frontier, budget, local) and policies without altering the retrieval architecture.

3. Key Contributions

1. Provenance-Grounded Model: A memory architecture that strictly separates immutable source evidence from canonical LLM-derived beliefs, enabling auditability and repair.
2. Two-Tier Write Path: A mechanism combining typed signal detection and hard-anchor validation to reduce unsupported extracted memories.
3. Deterministic Multi-Route Read Path: A retrieval system that operates without LLM calls, utilizing multiple routes (vector, keyword, temporal, graph) while preserving traceability to source evidence.
4. Model-Separable Interface: An architecture where memory delivery is decoupled from answer synthesis, allowing the same retrieved evidence to serve diverse answer models and policies.
5. Trace-Level Evaluation Protocol: A benchmarking approach that reports write-path, QA, and judge roles separately, preserving per-question artifacts (questions, gold answers, retrieved context, labels) for independent inspection.

4. Experimental Results

Eywa was evaluated on three benchmarks: LoCoMo (long-term conversational memory), LongMemEval-S (retrieval sufficiency), and BEAM (technical-memory stress).

• LoCoMo (C1–C4 Split): Using Claude Sonnet 4.6 for both write and QA roles with GPT-4o as the judge, Eywa achieved 90.19% judge accuracy.
  • Diagnostic runs showed that a stronger write path (Sonnet 4.6) significantly improved the performance of a smaller answer model (Qwen3 32B), raising its accuracy from 69.09% to 79.68%.
  • The system demonstrated robustness across different model families, though open-domain reasoning (C3) remained the most model-sensitive category.
• LongMemEval-S: Eywa achieved 88.2% retrieval-sufficiency accuracy across 500 questions. Failure analysis indicated that remaining errors were primarily due to retrieval coverage gaps and context-selection losses rather than simple extraction failures.
• BEAM: On a 700-question technical-memory stress test, Eywa achieved a mean nugget score of 81.45% and a pass@score≥0.5 of 85.29%. Performance was strongest in contradiction resolution, abstention, and temporal reasoning, while summarization and knowledge updates presented the most significant challenges.

Operational Metrics: The LLM-free read path demonstrated low latency, returning assembled context in roughly 150–200 ms on a stress store containing thousands of facts and links.

5. Significance and Claims

The paper positions Eywa not as a replacement for long-context models, but as a memory infrastructure layer that makes agent behavior inspectable and controllable.

• Auditability: By separating evidence from belief, Eywa allows users and developers to trace any claim back to its source turn, facilitating debugging, repair, and user-requested erasure.
• Diagnostic Clarity: The trace-level evaluation protocol transforms benchmark failures from opaque scores into specific engineering diagnoses (e.g., distinguishing between a retrieval gap and a synthesis gap).
• Cost and Privacy Flexibility: The architecture supports "local-first" deployment and model separation, allowing operators to choose between local privacy, low-cost budget models, or frontier reasoning capabilities without rebuilding the memory substrate.
• Refusal-Aware Evaluation: The paper highlights that standard metrics can be inflated by model refusals on answerable questions. Eywa's evaluation framework explicitly tracks refusal rates to provide a clearer picture of system capability.

The authors conclude that Eywa demonstrates that separating evidence capture, memory delivery, and answer policy creates a system that is both accurate and auditable, moving the field toward memory systems that are practical defaults for AI agents rather than opaque prompt decorations. All per-question artifacts and benchmarks are published for independent inspection.