OSATECHGPT.DEV
PL|EN
Architecture · Persistent Memory

AI agent memory runtime: persistent state without hidden memory

An agent memory layer should survive sessions without becoming an opaque database that silently changes behavior. Persistent state needs explicit structure, bounded execution and verifiable provenance.

CLAIM ≠ PROOF. No evidence = UNKNOWN.

Vector retrieval can help an agent find relevant text, but persistent agent memory has a broader job: preserve structured state, document relationships, activation history and write provenance across restarts and across multiple agent sessions.

NeurOSA Human Brain approaches memory as a programmable cognitive layer rather than a single embedding store.

.nsa → lexer → parser → AST → semantic analysis → type checks → IR 0.1 → runtime → activation → SQLite → hash-chain ledger

The runtime supports neurons, excitatory and inhibitory synapses, bounded impulse propagation, full activation traces, persistent state and restart recovery. The compiled IR defines the actual synapses through which impulses may propagate.

NeurOSA system pageSource repository

The native workspace supports Markdown documents, frontmatter, tags, headings, wikilinks, backlinks, pins, favorites, trash, immutable revisions and SQLite FTS5 search. The workspace belongs to NeurOSA and does not require Obsidian after migration.

A one-time read-only Obsidian importer can copy notes and safe attachments, generate neurons and synapses from wikilinks, and record import events in the ledger.

Connect adds persistent agent sessions, core context plus retrieval and atomic writes back into documents and the ledger. Ingest supports Drive material, PDF, DOCX, text files and chat exports. HTTP and MCP adapters use the same Brain API so persistence semantics do not diverge by transport.

Shared memory does not mean shared authority. The memory layer can retain state while execution authorization remains upstream in the governance layer.

The runtime exposes limits for hops, minimum impulse, event count and time, plus cycle detection and cancellation. Those limits prevent a memory activation graph from becoming an unbounded execution loop.

This is especially important for long-lived agent systems: persistence increases the amount of state available over time, so the activation policy must remain explicit and inspectable.

The append-only ledger is linked with SHA-256 hashes and stored with runtime state in SQLite. verifyLedger() detects changed payloads, deleted or reordered events, and invalid previous/event hashes.

EVENT n-1 HASH → EVENT n previousHash → EVENT n payload → EVENT n eventHash → verifyLedger()

The repository also states a recovery limitation: the current codebase is a functional reconstruction after loss of an older unpushed scratch workspace and does not claim byte-identical recovery of the missing source history.

NeurOSA Human BrainConcrete programmable memory runtime and ledger. Runtime GovernanceSeparates memory, authority, runtime and proof. Capability ControlControls which actions may mutate external systems. Agent Proof RuntimeVerifies execution outcomes independently.