Canonical memory
Original evidence remains addressable.
RetainedDevelopers & researchers
The machinery lives here. Inspect it.
Architecture, security, cryptographic evidence, benchmark methodology, and reproducibility are collected in one technical path for people who need more than the product story.
Technical map
Choose the depth you need.
The public product pages stay focused on outcomes. This section holds the system detail without diluting it.
ARCHITECTURE
HOM, AIMOS, and the trust boundary.
Responsibility boundaries, memory lifecycle, identity, retention, cognition, and native execution paths.
Open architecture ↗DOCUMENTATIONInstall, operate, secure, reproduce.
The release documentation contract, first-run ceremony, system operation, security, and benchmark reproduction.
Open documentation ↗RESEARCHQuestions, protocols, and papers.
The accessible research overview, current evaluation program, and publication commitments.
Open research ↗COMPATIBILITYProviders, local models, and tools.
The interoperability direction for model providers, local inference, IDEs, MCP, and agent protocols.
Open compatibility ↗Benchmark program
Three protocols. Complete denominators.
The release program separates utility measurement from adversarial retrieval security and keeps partial runs distinct from publication runs.
BenchmarkQuestionRun boundaryStatus
LongMemEvalLong-term interactive memoryFull 500-question runRunning
LoCoMoLong conversation memoryFull benchmark runQueued
PoisonedRAGTargeted retrieval poisoningPaired N=100 protocolPrepared
MutMem protocol
Mutable cognition with cryptographic traceability.
AIMOS allows cognitive weight to change as evidence changes while preserving the canonical memory and recording the transition. The paper examines the resulting chain of authorization, provenance, prior state, new state, and mutation evidence.
Cognitive state
Relevance can move in either direction.
MutableMutation proof
The state transition is independently verifiable.
LedgeredPublication evidence
A reproducible result is more than a score.
Each public run is expected to identify what ran, on which corpus, under which model roles, with which failures and cryptographic outputs.
Identity
Run ID, intended N, timing, hardware, and failure accounting.
Inputs
Dataset versions, manifests, hashes, and acquisition instructions.
Roles
Generator, retriever, judge, prompts, and relevant settings.
Artifacts
Receipts, memory hashes, mutation records, and verification roots.
Release-bound detail
No diagram substitutes for running code and evidence.
The full repository, installation ceremony, benchmark instructions, and evidence packs will be linked here when the release gates are complete.