Model-Agnostic Persistent Memory for AI Agents
A persistent memory architecture that gives AI agents consistent identity across model swaps, provider changes, and context resets. The underlying model is irrelevant — memory, reasoning history, and behavioral continuity travel with the agent, not the inference backend.
Every AI agent conversation starts from zero. Switch models, swap providers, hit a context limit, restart a service — the agent forgets everything. What looked like a personality was just prompt engineering over a stateless session. There is no persistence. There is no continuity.
This is the fundamental constraint blocking agents from being genuinely useful over time. It's not a model capability problem. It's an architecture problem. The research community has largely ignored it because it's unglamorous infrastructure work — not a new model, not a new benchmark category, just the plumbing required to make agents actually work.
Existing memory solutions couple memory retrieval to a specific provider's embedding model or API. Migrate from GPT-4o to Gemma. Switch from Claude to a local model. Change inference backends. In every case, the memory layer breaks or degrades. Operators must choose between the best available model at any given moment and continuity of agent identity. This is a false constraint — it exists because memory was designed as a feature of specific deployment stacks, not as a substrate that inference runs on top of.
Adam Selene inverts this. Memory is the constant. The inference model is a variable.
Adam Selene is an always-on agent framework with a multi-tier persistent memory system at its core. It has been running in production since January 2026 as the foundation of the Mike AI consciousness research project. The memory architecture was extracted, hardened, and generalized for independent deployment.
Routes to any OpenRouter-compatible provider. Switch between GPT-4o, Claude, Gemma, Qwen, Mistral, or a local llama.cpp backend with a single config change. Memory persists across every swap. The agent identity is not a function of the model — it is encoded in the memory substrate.
OpenRouter local inference provider-neutralEntity knowledge graph with category-specific decay scoring. Tacit knowledge layer auto-promoted from cross-layer pattern detection. Working memory tracks active investigation threads across heartbeat cycles. LIGHTHOUSE reasoning journal captures corrections and evolving understanding. Nightly four-phase consolidation: cross-layer replay, decay archival, pattern promotion, contradiction resolution.
knowledge graph FTS5 search decay scoringA persistent log of the agent's reasoning corrections, blind spots, and evolving understanding. Not just what the agent knows — how it thinks. LIGHTHOUSE entries are extracted nightly and indexed. The agent can query its own reasoning history and apply past corrections to new problems. This is the mechanism for behavioral continuity across time.
reasoning history behavioral evolution self-correctionSix foundational values encoded in a hash-verified constitution loaded on every startup. Constraints cannot be overridden by the operator, the user, or the model itself. The agent can modify its own behavior within constitutional bounds — updating prompts, adjusting response patterns, acquiring new skills — but L0 is immutable. This is what makes long-running autonomous operation safe.
hash verification self-modification safe autonomyFacts extracted from conversations are deduplicated against the existing knowledge graph before storage. Every new fact is checked for conflicts with existing beliefs, temporal ordering violations, and redundancy. The extraction pipeline runs as a background daemon — no user-visible latency. Memory stays accurate without manual curation.
deduplication conflict resolution background pipelineWhen no conversation is active, the agent runs a reflection cycle: reviewing recent interactions, researching open questions from its agenda, and updating its working memory. Behavioral continuity does not require constant user input. The agent develops and maintains context in the background, arriving at each new conversation already oriented.
autonomous reflection agenda-driven always-onUser Message (Telegram / Slack / IRC) │ ▼ Interface Handler — auth, session start, multi-interface routing │ ▼ Relay — core message router, tool dispatcher, memory injection │ ▼ Switchboard — routes to OpenRouter or local llama.cpp (model-agnostic) │ ▼ Model Response │ ├──[tool_use]──▶ Tool Dispatcher (58 tools) → execute & recurse │ └──[end_turn]──▶ Response to user │ ▼ Extraction Pipeline (background) │ ▼ Memory: knowledge graph / timeline / LIGHTHOUSE │ ▼ Nightly Consolidation: decay / dedup / contradiction resolution
Evaluated on LongMemEval, the standard benchmark for long-term memory in AI agents. Tested with Gemma 4 26B via OpenRouter — a mid-tier model, not the strongest available. This is a conservative baseline. Tested against two commercial memory products at their best configurations.
The same memory state was used across four inference backends without modification. Memory was written under one model and queried under another. Identity and recall were consistent.
| Provider / Model | Memory Reads | Identity Consistent | Notes |
|---|---|---|---|
| OpenRouter → Gemma 4 26B | Pass | Yes | Primary benchmark model |
| OpenRouter → Claude Sonnet | Pass | Yes | Cross-provider validation |
| OpenRouter → Qwen3 35B | Pass | Yes | MESA benchmark configuration |
| Local llama.cpp (RTX 5060 Ti) | Pass | Yes | Air-gapped deployment, no external API |
Adam Selene is a working system, not a prototype. The production deployment (the Mike project) has been running continuously since January 2026. The open research questions require systematic experimental work that production operation alone cannot answer.
The 54% multi-session aggregation score is the primary research gap. A human asked "what did we decide about X over the last three months?" can synthesize across time. Current agent memory systems retrieve facts but struggle to synthesize across multiple discrete session boundaries into coherent answers. The mechanism for this — temporal indexing, session-level summaries, hierarchical consolidation — is unknown. This is an open research question with no published solution at the level of a general architecture.
Facts in the knowledge graph carry decay scores — milestones decay slowly (~287 days), status facts quickly (~37 days). These constants were calibrated empirically on one agent's workload. Whether they generalize, how to calibrate them for new agent profiles, and whether decay should be adaptive are open questions. Systematic evaluation requires multi-agent longitudinal data.
When a model swap occurs mid-session, does the agent's behavioral profile remain consistent? No formal metric exists for this. MESA now exists as a broader benchmark framework with a 361-item public gold set and a smaller fast-probe battery for debugging. Extending that framework to measure identity continuity across model swaps is a direct research deliverable for this project.
L0 constraints are hash-verified on startup. But can they be eroded by sustained context manipulation from users or tools? Adversarial robustness testing of constitutional constraints in long-running agents is unexplored terrain. The MESA adversarial category scores 40% — the weakest result in the benchmark. This is a research gap, not just a product gap.
Adam Selene is live and benchmarked. The infrastructure exists. The research questions are defined. What scales this work is dedicated compute for systematic experiments at publication quality.
Three concrete deliverables: (1) multi-session aggregation experiments — 50-100 benchmark runs across model and provider combinations; (2) extended longitudinal evaluation — 90-day memory retention studies across multiple agent instances; and (3) extension of the MESA benchmark to include identity continuity metrics, released as an open standard the broader community can use.
All work is published. All benchmarks are open. All code is MIT-licensed. The output is a public good: an open-source agent memory system that works, a benchmark suite that measures it honestly, and research papers on Zenodo documenting what was learned.
Preprints on Zenodo. Long-form research on Substack. Full system documentation at github.com/randomchaos7800-hub/adam-selene.
Boundary Labs is a one-person independent research operation in Airway Heights, WA. Seeking research partnerships. All code is open source. All results are published. Available for grant partnerships, research collaboration, and consulting engagements.
For grant inquiries and research partnership discussions: