Imagine that a historical conversation contains two statements at different times: 'Yutou gave a bag of biscuits to Tutu' and 'Yutou works at a court.' When the user asks, 'Where does the person who gave the biscuits work?', the system cannot answer from a single keyword match. It must first establish that Yutou gave the biscuits, then find Yutou's workplace, and provide historical evidence supporting both conclusions.

Such questions illustrate three fundamental requirements for long-term memory systems. First, untraceable summaries must not completely replace original conversations. Second, retrieval must account for both original-text details and structured representations. Third, the model's role should be limited to assessing and ranking candidate evidence, rather than freely generating new memories.

During writing, the conversation's original text is first persisted as RAW records. In model mode, the system then extracts clearly bounded FACT expressions supported by the original text. FACT does not replace RAW; it is bound to its source message through source_message_id. Facts facilitate retrieval, while the original text preserves context and grounds for verification.

• RAW preserves personal relationships, original wording, temporal clues, and contextual details.

• FACT provides more compact, retrieval-friendly expressions of facts.

• source_message_id allows every fact to be traced back to its supporting origin, preventing memory from becoming an isolated summary.

A Search request carries an exact user_id constraint from the outset. The system performs FTS5/BM25 retrieval separately on the raw-message and fact indexes: the former preserves relationships and original context, while the latter directly matches concise, explicit factual expressions. Candidates from both paths are fused and deduplicated into an evidence pool for the ranking module.

• The raw-message channel compensates for contextual compression introduced by fact extraction.

• The fact channel makes retrieval more direct for questions that span different phrasings and dialogue turns.

• Dual-path candidates prevent a single pathway from determining all memory recall results.

In the officially evaluated version, gpt-4o-mini assesses and ranks candidates that have already been retrieved. It receives the current query and candidate set, and can return only ordered_ids that actually exist in that set. The system then revalidates whether the candidate IDs exist, belong to the current user, and, for facts, retain valid provenance. The model neither generates the final answer directly nor creates new memory entries.

Design Boundary: The model chooses among existing evidence; the storage layer guarantees that evidence is authentic, isolated, and traceable.

ChronoHybridMem's goal is not simply to increase the number of memory entries, but to shorten the distance between the current answer and its historical basis. In the example above, the returned evidence includes the raw message identifying the biscuit giver, the structured fact describing the workplace, and its source link. A downstream Agent can answer from this evidence and return to the original text for verification when needed.

This mechanism translates reliability requirements into engineering boundaries. Idempotent writes reduce the risk of duplicate-memory contamination; SQLite WAL supports stable persistent reads and writes; both API and SQL queries use user_id as the isolation boundary. If configuration required for model mode is missing, the system fails explicitly rather than silently switching to unknown behavior.

The Rank 5 and Overall 44.33 stated at the beginning refer to v0.2.0. All P1–P5 results discussed below are local research completed after official evaluation, mainly using public LoCoMo data and a local Qwen3-4B proxy model. They compare retrieval mechanisms, validate engineering boundaries, and screen research hypotheses; they are not new official leaderboard scores.

We use a layered evidence-reporting framework. Official results describe historical performance verified by the platform. Full local runs on 1,977 questions compare method changes under the same proxy model. Fixed sets of 20, 100, and 200 questions, and synthetic strata, are used only for mechanical checks, diagnosis, and promotion gates. We discuss method gains only after paired comparisons using the same data, model, query plans, and call budget.

research-v0.3.0 and research-v0.4.0 are also post-challenge local research milestones. Both were evaluated under a strict evidence-turn matching protocol on 1,977 LoCoMo questions. Neither is a new official submission, and neither changes the fact that v0.2.0 achieved Rank 5 and Overall 44.33.

All post-challenge numbers are labeled local proxy results. Methods that fail predefined thresholds retain their code and experimental records but do not enter the default path.

Post-challenge development was not simply a matter of stacking larger models or more retrieval channels. We formulated each stage as a falsifiable question: what failure type does it address, does it add model calls, can it preserve source traceability, and can it be disabled when performance regresses? This produced the successive P1–P5 stages.

Before the structured-retrieval research in P1–P5, the project passed through two local-method milestones, research-v0.3.0 and research-v0.4.0. They validated hybrid recall and dedicated reranking respectively, laying the experimental groundwork for further separating query planning, recall, fusion, and ranking.

v0.3.0 retained the original Add/Search API and raw-evidence return contract, while using Porter-normalized BM25 and BAAI/bge-large-en-v1.5 dense retrieval in parallel during Search. Each path's candidate scores were z-score normalized within its own candidate set and fused with alpha=0.3. Only the Top-5 fused results were then reranked with answerai-colbert-small-v1 using MaxSim late interaction. The entire pipeline reranks only existing candidates, ultimately returns raw messages, and does not generate answers.

In the strict evidence-turn matching evaluation on 1,977 LoCoMo questions, v0.3.0 achieved Hit@1 0.4355, Hit@3 0.6186, Hit@10 0.7577, and MRR 0.5183. Compared with the model-free lexical baseline's Hit@1 0.3359 under the same protocol, this was an improvement of 0.0996. Pool-size experiments on a fixed 200-question set also showed that ColBERT Top-5 outperformed Top-10 and Top-20. A larger reranking pool could improve deeper recall but hurt first-place ranking, so the final configuration was frozen at Top-5.

v0.4.0 shifted the main focus from candidate recall to top-ranked ordering. On v0.3.0's lexical–dense candidate pool, it used the first-token yes/no probabilities of local Qwen3-Reranker-4B to assess query–evidence relevance for each candidate and rank the Top-10. The model ran through a loopback OpenAI-compatible service and output only relevance probabilities and an ordering of existing evidence. Model weights, data, and credentials were not included in the repository or Docker image.

This stage also added time-aware dense keys: message dates were incorporated into retrieval representations as controlled metadata, with a fixed weight of 0.5. Used alone, time keys did not improve Hit@1 on the full dataset, but expanded Top-10 candidate coverage. Combined with dedicated Qwen reranking, this additional recall translated into small, stable improvements near the top of the ranking. Ultimately, v0.4.0 achieved Hit@1 0.5225, Hit@3 0.6808, Hit@10 0.7653, and MRR 0.5856 on 1,977 questions. Relative to v0.3.0, Hit@1 increased by 0.0870 and MRR by 0.0673.

P1 decomposes a single natural-language query into intent, core terms, expansions, entities, temporal cues, and up to four evidence needs. Different fields enter existing lexical channels separately, while evidence needs serve as bounded query-side signals reusable in later experiments. P1 adds no Search-stage model calls and does not change the Add/Search API. When the planning output is incomplete, the system falls back to safe single-stage retrieval.

In local screening on a fixed 200-question set, the structured planner improved Hit@1 from 0.545 to 0.565 and MRR from 0.6145 to 0.6292 compared with the flat planner, while Hit@10 remained 0.740. In the subsequent full local run on 1,977 questions, P1 achieved Hit@1 0.5761, Hit@3 0.7157, Hit@10 0.7618, and MRR 0.6479.

Multi-hop questions often require multiple complementary pieces of evidence, but ordinary ranking can place similarly worded records with duplicate sources near the top together. P2 therefore attempted deterministic selection among P1's existing candidates according to the additional coverage of evidence needs. It created no new candidates, changed no candidate content, and added no model calls. On a fixed 20-question set, P2 matched P1's Hit@1 and improved Hit@3 and MRR.

However, on 35 public synthetic stratified cases with frozen query plans, Hit@1 dropped from 1.00 to 0.8571 and MRR from 1.00 to 0.9286. The reason was that covering more need-related terms is not equivalent to being more suitable as the first piece of evidence. P2 was therefore explicitly rejected from the default path and retained only for reproducible experiments and failure analysis. This result shifted our focus from diversity near the top to precise attribution of recall failures.

P3 explored representing people, locations, organizations, relationships, and temporal updates with an evidence graph, while maintaining one constraint: every node, entity mention, and relationship edge must be independently witnessed by a raw message, and search must ultimately return only raw evidence. This avoids treating unverifiable model-extracted relationships as real memories.

The strict relationship graph P3-A produced only 3 independently witnessed relationship edges from 419 raw messages in a fixed slice, showing how sparse graph structure becomes under high-precision constraints. P3-B1 further built source-local entity-mention anchors, covering 363/419 messages (86.63%), but on a fixed 20-question set Hit@1 fell from 0.40 to 0.35 and Hit@10 from 0.55 to 0.50. P3 therefore did not become a global default retrieval channel. Its central lesson was that edge count or coverage cannot establish a graph's value; that value must be demonstrated jointly through paired retrieval gains and provenance audits.

P4 first established per-question recall auditing. Among a fixed 100 questions, there were 30 Top-10 recall failures: 20 were fusion misses, where relevant evidence had been found by at least one channel but failed to enter the fused candidate pool; 8 were channel misses, where no lexical channel found it; and only 2 were reranker drops. This distribution indicated that the main bottleneck was not the model misranking correct candidates, but evidence failing to enter the reranking pool reliably.

To address the largest fusion-miss category, P4-A treated each evidence need already generated by P1 as an independent, bounded retrieval channel, reserved 2 candidate slots for each, and passed those candidates together with the others into the original reranking pipeline. The method reused the same query planning, added no Search model calls, and generated no new evidence.

P4-A's per-question diagnosis showed that candidate swaps could produce Top-1 hits for some questions, but could also displace originally correct Top-1 results for others. P5 therefore attempted swaps only when Top-1 and Top-2 were close and the second candidate had stronger evidence signals. We tested channel counts, query-token overlap, temporal/correction strata, and model-reported confidence in sequence.

None of the four signal types passed the fixed 200-question gate. Channel counts changed Hit@1 by -0.005; query overlap by -0.035; narrowed strata by -0.010. Confidence gating never triggered, while the prompt asking the local model to report confidence itself lowered Hit@1 by 0.045. The conclusion was that, without gold-answer leakage, these simple signals could not reliably determine when to swap candidates. P5 remains as disabled-by-default ablation code, and further tuning of gated swaps was stopped.

The challenge rules have been updated, so the first priority for the next stage is to turn the new rules into an executable, auditable protocol, then determine which modules are suitable for a candidate version. The technical roadmap continues to follow four principles: traceable provenance, bounded candidates, paired experiments, and fallback on failure.

Roadmap Summary: In the near term, freeze the evaluation protocol under the new rules and the recall_miss diagnostic set. In the medium term, validate source-constrained adjacency, bridging, and temporal recall. In the long term, establish a model-independent evaluation framework covering effectiveness, cost, provenance integrity, and stability.

Facts in long-running conversations change: users may revise times, cancel plans, correct names, or add applicability conditions. The next stage plans to model state changes as evidence chains of original statement → update/correction → currently valid state, while preserving all historical sources.

When the system outputs a current state, it must explain which old information it supersedes and on what grounds, rather than retaining only the latest summary.

The key to long-term memory is not compressing historical conversations as much as possible, but enabling the system to find correct, complete, verifiable evidence when recall is actually needed. Raw messages, source-bound facts, dual-path candidate retrieval, and constrained ranking are not independent features: together they form a closed loop from writing and recall to evidence return.

The P1–P5 results also yield three interim lessons. First, stable gains from query planning may matter more than blindly adding retrievers. Second, high coverage, structural complexity, or strong explainability does not automatically mean good retrieval; paired experiments must validate it. Third, negative results are not the end of development, but evidence that helps a system narrow itself to clearer boundaries. Precisely because some hypotheses in P2, P3, and P5 were rejected, P4 could concentrate on the recall failures that truly dominated.

We believe long-term memory for Agents should treat traceable factual representations, preservation of original sources, candidate-scope constraints, precise user isolation, and falsifiable evaluation protocols as foundational capabilities. Only then can memory systems progress from retrieving text that seems relevant to an answer toward evidence systems capable of supporting complex tasks.