简体中文 | English
Relata is an open, community-grounded frontier memory research lab. It defines, synthesizes and critically examines agent memory, open-source architectures and the evidence for their strengths and limitations, inside and outside long-term human–AI relationships. Chinese is the primary R0 working and community language; the English entrypoint is maintained alongside it.
What does it mean for an agent to remember, how do different systems implement it, and what enables continuity across changing tasks, lives and relationships?
Relata brings together its own benchmark research, a collection of system studies, and an incubator for future memory systems. ADR-0008 makes these mutually supporting functions core to the project.
| Function | What Relata develops | Start here |
|---|---|---|
| Own benchmark | Questions, cases, controls, evaluation methods, experiments and evidence-backed comparisons | Case Lab, current evidence |
| System collection | Inspectable memory architectures, source studies and conditional comparisons; museum / sample room / lab remains an open naming choice | System Census, ten source studies, architecture atlas |
| Incubator for future systems | Design hypotheses, alternatives, counterexamples and lessons that can inform independent systems such as Tilia | Candidate design, case-to-test map |
System studies can expose case questions; evaluation can challenge architectural claims; both can inform future designs. Each function also has independent research value. The benchmark is under development, the existing collection contains research drafts, and future Tilia product source belongs in its own project.
Agent Memory Study (AMS) contributes readings, source investigations and experiments that Relata can use with their original evidence limits. Both projects can support long-form synthesis while retaining their own questions and artifacts.
General agent memory and adult long-term intimacy are both research contexts. Intimacy and romance remain a founding focus; the existing Case Lab studies ordinary personal life, shared relational experiences, operational/project decisions and artifacts, and companion/system continuity as a longitudinal mixed-domain ecology. Ordinary events do not need invented romantic symbolism to count.
ADR-0006 establishes this broader identity. Definition, architecture studies, comparative essays and criticism are research outputs alongside cases. The agent-memory inquiry connects a standalone article direction, selected public AMS research and ten bounded source studies. The architecture atlas adds thirty source-linked SVG views and an offline interactive reader. These are research drafts, not an exhaustive survey or comparative performance results.
Relata asks what is retained, activated, admitted to context, used, repaired, or appropriately left silent. Factual recall, temporal reasoning, source fidelity, noise resistance, scope isolation and full-history/full-search controls remain valuable; their observation boundaries must be explicit.
The research may inform Tilia, a future independently authored memory system starting from an empty repository. Tilia is a separate long-term goal; Relata remains the research lab. The candidate design and case-to-test map turn existing studies into choices, controls and counterexamples, without settling Tilia's architecture. Its intended design requires vector-based semantic retrieval; encoder/model choices remain open.
A very early explicit-input experiment already exists in a separate repository under the historical name relata-memory, with its own private remote. Its code is not publicly published and currently offers literal search with a scripted consumer. The corrected ADR-0007 records that implementation began from an overreading of the research goal; retaining the experiment does not make it formal Tilia or an ongoing build mandate. Its engineering record preserves bounded checks, not semantic memory outcomes. The lab’s other research outputs remain independently valuable.
R0 — Research Foundation. Two narrow source Evidence Cards are accepted. RC-001 is clinic-ready; RC-002/003/004 are unreviewed seeds. There are no reviewed System Cards, accepted cases, validated evaluators, semantic system-evaluation results or rankings.
Software now includes the public repository checker and an offline RC-002 execution rehearsal. It runs two synthetic histories across five conditions with a deliberately scripted subject, preserving evidence and exporting a blind packet. Its tests establish plumbing behavior only. No model is called, and no semantic or capability score is generated. Exact evidence status is in STATUS.
A second offline tool prepares and verifies 18 unanswered inputs for RC-005 / shared-work authorship and audits whether input projections erase the speaker distinction. The case remains an authored candidate with no independent human review or system evaluation. See the input audit guide.
The proposed first research cycle connects a case portfolio, four candidate system configurations, independently deliverable source/runtime exhibits and mechanism tests for future designs. ADR-0009 proposes a bounded exploratory execution route. The RC-005 → Mem0 execution packet specifies the proposed reader, controls, timed/recovery observations and cost cap; native runnability remains unverified. These are planning artifacts: no live study, approved spending, accepted benchmark result or hosted research site follows from them.
A follow-up correction-race challenge adds authored interleavings, counterfactual histories and recovery controls; a pinned Hindsight source audit examines existing safeguards. These remain draft research, with no system execution or defect claim.
Three bounded research additions connect these functions: an unreviewed RC-006 companion-migration seed, a correction-propagation comparison based on existing pinned source studies, and a multilingual semantic-retrieval study. They develop cases, competing explanations and research choices; they add no accepted case, system result or selected Tilia architecture.
No canonical ontology, system protocol, scoring contract, benchmark release, Leaderboard, Arena, SDK, service or hosted infrastructure is accepted. The historical architecture draft remains non-normative; the assumption register preserves its disposition.
ADR-0005 authorizes replaceable offline pilot tooling without waiting for third-party scores. It does not authorize provider adapters, private data or public performance claims. Formal promotion gates still apply to accepted cross-system boundaries.
EC-001 addresses AML's public boundary and causal limits. EC-002 addresses PM-Bench observation/scorer binding and records no observed step-order impact on its released corpus. Neither adopts or validates the source benchmark.
Start the local rehearsal now (Python 3.10+, standard library only). Use a new output directory on every run; its parent must already exist:
python3 -B tools/synthetic_pilot.py run --output experiments/artifacts/rc002-demo
python3 -B tools/synthetic_pilot.py verify experiments/artifacts/rc002-demoThe rehearsal guide explains the five conditions, subject/evaluator views, review packet, integrity checks, errors, and limits. RC-001 human calibration, seed refinement and public-source System Cards can proceed in parallel. The broader execution path remains research guidance; the offline exception does not make it an accepted protocol.
Community members participate as co-researchers. Useful contributions include exact-source Evidence Cards, abstract Incident Seeds, system-native System Cards, case review, and governance critique. No raw-chat contribution is required.
Consent is per contribution, not a ladder: read consent modes and participation principles. Restricted consent-record stewardship must be settled before sensitive collection.
Public visibility allows inspection; it does not authorize raw private conversations, identifying consent records, credentials, restricted system details or private review material in this repository. The CLI runs bundled invented adult material only. The checker reads Git's public working set and does not recursively inspect ignored local experiment records.
Public licenses remain an open decision. No formal release or accepted result publication is implied by this source branch.
Charter, research questions, status, assumptions, terminology, language policy, source research, systems, cases, community, governance, decisions, and vision history.
Follow STATUS, CHARTER, ASSUMPTION_REGISTER, accepted decisions, research/governance/case materials, then non-normative vision history. Run from the repository root:
python3 -B tools/check_repo.py
python3 -B -m unittest discover -s tests -vChecks validate public-file structure, links, bilingual declarations and paired changes, selected case markers, and isolated execution/evidence regressions. They do not establish semantic translation equivalence, reviewer agreement, memory necessity, long-term retention or system quality.