Research
Evidence for governed execution memory.
Benchmark reports, reruns, and technical notes behind Aionis: context compression, admission policy, external-agent continuation, and memory firewall behavior.
Evidence Hub
Four claims with data behind them.
The strongest Aionis numbers are about preserving executable state while reducing prompt mass and keeping memory influence auditable.
Recall Engine
40/40product-positive casesStrict ID-neutral active-state recovery
MGBench strict reruns with optional Zvec ANN: 100% active recovery, 0 unsafe direct-use, 100% rehydrate recall, and 100% trace coverage.
- fresh Runtime
- fresh Zvec sidecar
- 1,305 avg context chars
Context Compression
77.2%mean compressionState-preserving context compression
100-scenario suite: 100% current-state, negative-memory, and procedure retention with 0% stale and forbidden leak.
- 610.95 mean chars
- 100% audit
- 95.8% LLM action accuracy
External Agent E2E
219kAionis prompt tokensFull-history-level continuation with less prompt mass
40-record five-arm run: same accepted-direction outcome as full history with materially lower prompt cost.
- 985k full history
- 344k BM25
- 624k Mem0
Admission Policy
776labeled rowsMemory influence becomes measurable
Admission dataset records what was surfaced, suppressed, used, and followed by which outcome.
- 55 task signatures
- 0 hard-boundary direct use
- real LLM rerun
Memory Firewall
0%unsafe direct-use after firewallGovern external memory candidates
Local Mem0 A/B: same retrieval candidates, Aionis governs admission before prompt influence.
- 83.3% raw Mem0
- 100% primary route
- 100% audit
Open Benchmark
MGBench: memory governance as a public test surface.
MGBench is an open benchmark for evaluating whether memory systems preserve useful long-term context while blocking stale, invalidated, cross-scope, contradictory, or low-authority historical memory from becoming agent-usable context.
0.1 manifest suites
Credibility governance
Ordinary-memory governance
Controlled forgetting
High-trust conflict governance
Scope isolation
Lifecycle inference
Execution-tree effect
Execution-tree stress
Benchmark map
Use these as the Research page backbone. Older DT artifacts stay as appendix material.
Candidate retrieval engine
MGBench strict ID-neutral 40-scenario runs with optional Zvec ANN
40/40 active recovery, 0/40 unsafe direct-use, reproduced on a fresh RuntimeState-preserving compression
100 deterministic scenarios + 24 LLM-scored downstream trials
77.2% compression, 0% stale/forbidden leak, 95.8% downstream accuracyExternal agent continuation
Five-arm continuation runs plus an isolated two-session Claude Code host case
Governed context carries active execution state across session boundaries with materially lower prompt massAdmission policy flywheel
776 admission rows, 55 task signatures, shadow and real-agent reruns
Memory decisions are exportable, comparable, and replayableExternal memory governance
Local Mem0 A/B and ordinary-memory horizontal evals
Aionis controls direct-use leakage and preserves audit coverageResearch notes
Articles that explain the runtime decisions behind the benchmark results.
Execution memory is not chat history
Long-running agents do not only need recall. They need a durable state model that tells the next run which route survived, which alternatives were retired, and which evidence must stay auditable.
Read noteMemory Firewall: admission before influence
A Memory Firewall sits between recall and the agent prompt. It prevents relevant-but-unsafe memory from acting as instruction.
Read noteStrong retrieval below. Strict governance above.
The optional Zvec ANN backend improves candidate generation without changing Aionis' core governance contract: retrieval proposes, admission decides.
Read noteState-preserving context compression
Aionis does not optimize for the shortest possible summary. It optimizes for compact context that preserves current state, non-current alternatives, reusable procedure, rehydrate pointers, and an audit trail.
Read noteThe admission policy flywheel
Aionis records which memory was surfaced, suppressed, used, and followed by which outcome. That creates the dataset for measuring memory admission quality.
Read note