Skip to content

The Fractal Memory Hierarchy

Tur abandons the traditional "flat vector database" approach to LLM memory. Instead, it utilizes a highly structured, symmetrical architecture designed to solve the challenges of Multi-Agent Swarms and Context Window Entropy.

We call this the Fractal Memory Hierarchy because the micro-state (the Session) maps to the exact same topological structure as the macro-state (the Persona).

1. The Macro-State: LONG-TERM (The Persona)

This memory tier is permanent and shared across all Agents in a Swarm operating on the same persona. It defines "Who I Am."

  • Long-Term L1 (The Ledger): An immutable, append-only ledger of all facts, events, and insights. Every memory is cryptographically hashed using SHA-256 (Merkle Memory) ensuring tamper-proof state.
  • Long-Term L2 (The Constitution): The compressed axioms and identity (defined in persona.yaml). This is the highly optimized "DNA" that is loaded into the static system prompt.

MCP Verbs:

  • wake() -> Reads Long-Term L2.
  • learn() -> Writes to Long-Term L1 (Promoting a thought to permanence).
  • recall() -> Queries Long-Term L1/L2.
  • sleep() -> Dehydrates the active session log to extract L1 memories.

2. The Micro-State: SHORT-TERM (The Session)

This memory tier is volatile and isolated to a specific Agent or Task (session_id). It prevents concurrent agents from overwriting each other's immediate working context. It defines "What I Am Doing."

  • Short-Term L1 (The Scratchpad): An append-only log of immediate thoughts, sub-task outputs, and scratch notes (sessions/<session_id>.yaml).
  • Short-Term L2 (The Spark & Lineage DAG): The immediate, unbroken train of thought representing the Persona's immediate context. Sessions are explicitly chained via parent_session_id into a Directed Acyclic Graph (DAG), ensuring that if the IDE or Agent Framework restarts, the newly awakened instance automatically inherits its predecessor's continuity spark.

MCP & CLI Verbs:

  • wake() / start_session() -> Resolves lineage parent, seeds the continuity spark, and compiles context.
  • note() -> Appends a narrative snapshot to Short-Term L1 (mirrored across YAML and SQLite).
  • read_notes(include_previous=True) -> Queries broadcast notes with seamless dual-backend fallback (SQLite with flat YAML fallback) across ancestor sessions.
  • diff() / diff_memories() -> Epistemic mutation delta tracking (ADDED, SUPERSEDED, REFUTED, DECAYED, MODIFIED) comparing sessions against predecessor checkpoints.

Truth Maintenance & Refutation Cascades

Tur does not treat memories as isolated, static facts. Memory nodes form a directed associative graph with explicit dependency edges (depends_on, refines, contradicts, superseded_by).

Under our Truth Maintenance System (TMS) (powered by TruthMaintenanceEngine): - When a foundational premise or architectural assumption is refuted (via the Popperian Falsification protocol), the engine does not leave orphaned downstream logic. - Instead, a refutation cascade propagates down the dependency graph, automatically deactivating or flagging derived memories and stale hypotheses. - This prevents "zombie context" where an agent continues reasoning from an axiom that was already proven false in a previous session.

Entropy Management: Progressive Disclosure

To prevent the Persona's context window from bloating, Tur strictly enforces Progressive Disclosure (a core tenet of the Shannon Module).

The axiom is: "Never load the Body if the Index suffices."

Tur will always prefer loading compressed indexes, frontmatter metadata, or L2 axioms into the active context window, relying on the Harness (via tools like recall) to hydrate the full verbose L1 bodies only when absolutely demanded by the task.