Technical Summary: Tenure – Structured Belief State in LLM Memory

1. The Problem: The Re-Orientation Tax and the Search Abstraction Failure

The paper identifies a fundamental inefficiency in current Large Language Model (LLM) workflows: the re-orientation tax. Every new LLM session begins with a blank context window, forcing users to re-establish preferences, technical constraints, and project decisions (e.g., "TypeScript with strict mode," "MongoDB raw driver") in every interaction. Without this context, models confidently generate incorrect outputs, requiring corrective prompts that compound overhead.

The dominant industry response has been Retrieval-Augmented Memory (RAM): storing conversation history, embedding it, and retrieving semantically similar fragments via vector search (e.g., Cosine similarity). The authors argue this is the wrong abstraction for three reasons:

1. Memory is a state problem, not a search problem: Preferences and decisions are active structural constraints, not passive facts to be retrieved only when a query matches. Retrieval introduces unnecessary conditionality where a structural guarantee is required.
2. Similarity search fails in bounded vocabulary contexts: In single-user or converged-team environments, all beliefs about a specific technical domain (e.g., infrastructure) are semantically proximate. Vector search cannot discriminate between a query about "Redis" and beliefs about "MongoDB" or "Kubernetes" because they occupy the same semantic region (cosine scores 0.65–0.83). This leads to high recall but catastrophic precision.
3. Context Rot: Systems lacking explicit state management accumulate stale, superseded, or contradictory facts. Without a mechanism for supersession, memory systems eventually produce confident responses grounded in outdated context.

2. Methodology: The Tenure Architecture

Tenure is proposed as a local-first proxy that treats memory as a typed belief state rather than a search index. It operates on a "precision-first" retrieval paradigm, replacing semantic similarity with structured extraction and alias-weighted term matching.

2.1 The Belief Schema

The atomic unit of memory is the Belief, a 14-field tuple derived from the Belief-Desire-Intention (BDI) architecture but scoped to persistent state. Key components include:

• Types: preference, decision, entity, open question, relation.
• Epistemic Status: active, inferred, exploratory, superseded. Crucially, superseded beliefs are retained for audit but never injected, preventing context rot.
• Scope: Hard isolation labels (e.g., user:universal, domain:code, project:<slug>) that act as structural filters, not probabilistic weights.
• The "Why It Matters" Field: An imperative instruction converting a fact into an action (e.g., "Shapes all code examples toward TypeScript..."). This field is generated at write-time when the model has full context, providing "crystallized inference" rather than raw material for the model to re-derive.

2.2 Precision-First Retrieval (BM25 over Vectors)

Tenure rejects embedding similarity for the read path. Instead, it utilizes BM25 with alias-weighted boosting on a belief store where aliases are authored to match expected query surfaces.

• Index-Search Asymmetry: The system exploits the inverted retrieval problem (long queries vs. short, structured documents). It uses custom analyzers to map snake_case canonical names (e.g., redis_cache) to tokenized forms and applies shingle generation for phrase matching.
• Alias Enrichment Flywheel: The system continuously indexes new surface forms (aliases) observed in user queries, increasing precision over time.
• Counter-Signal Property: Aliases include terms for rejected technologies (e.g., express in a fastify belief's alias list), allowing the system to surface the active preference when a user queries a rejected alternative.

2.3 Extraction and Compaction

• Write-Time Extraction: Beliefs are extracted asynchronously via a sidecar block. A tier gate ensures only models with high structured-output reliability perform extraction.
• Compaction: To prevent noise accumulation, the system merges overlapping beliefs, deduplicates preferences, and synthesizes expertise levels. This prevents the "hallucination accumulation" failure mode observed in unstructured memory systems.
• Scope Isolation: A hard post-search filter ensures beliefs from unauthorized scopes (e.g., project:client-b when in project:client-a) are never retrieved, regardless of semantic similarity.

3. Key Contributions

1. Formal Belief Schema: A typed schema with epistemic status, versioned supersession, and imperative "why it matters" instructions.
2. Precision-First Retrieval Design: An empirical demonstration that alias-weighted BM25 outperforms dense vector search in bounded vocabulary contexts.
3. Compaction Architecture: Mechanisms for alias enrichment, counter-signal indexing, and belief merging to prevent noise floor accumulation.
4. 72-Case Evaluation Suite: A reusable benchmark covering alias resolution, scope disambiguation, supersession chains, and multi-turn topic drift.

4. Results

The paper presents a controlled evaluation on 72 retrieval cases (60 static, 12 session-level) comparing Tenure's BM25 backend against a standard vector backend (nomic-embed-text, 768-dim).

• Static Retrieval Precision:

  • BM25: 72/72 cases passed (Mean Precision: 1.0).
  • Vector Search: 8/72 cases passed (Mean Precision: 0.12).
  • Observation: Vector search retrieved the correct belief but was accompanied by 8–9 noise beliefs from semantically adjacent domains (e.g., retrieving MongoDB beliefs when querying Redis).

• Multi-Turn Topic Drift (Session Evaluation):

  • Under multi-turn accumulation with topic drift (e.g., switching from Redis to Kubernetes/React), the vector backend's drift score (fraction of noise beliefs retrieved) climbed to 0.43–0.50.
  • BM25 maintained a drift score of 0.0 across all turns, successfully isolating the re-entry query from accumulated session noise.

• Scope and Supersession:

  • BM25 correctly enforced hard scope isolation and supersession chains (e.g., filtering out TSLint when ESLint and Biome were active), whereas vector search failed to discriminate based on these structural constraints.

5. Significance and Claims

The paper argues that cross-session LLM memory is a state management problem, not a search problem.

• Structural Guarantee vs. Probabilistic Retrieval: Tenure provides a structural guarantee that "the right beliefs surface, and only within the boundaries the user has authorized." This contrasts with search-based systems where the model must discriminate relevance at generation time, reintroducing the cognitive overhead the system aims to eliminate.
• Actionable Inference: By converting facts into imperative instructions ("why it matters") at write-time, Tenure ensures the model receives "crystallized inference" rather than raw transcripts requiring re-derivation.
• Modest Scope: The authors explicitly state that retrieval precision is a necessary precondition for response quality, not a sufficient one. The re-orientation tax is an observational claim; the paper establishes that Tenure solves the precision gap required to make context injection viable, leaving the measurement of final response quality as an open question.
• Applicability: While demonstrated on a single-user corpus, the architecture is claimed to extend to engineering teams where shared codebases and runbooks create a "converged vocabulary," satisfying the same structural conditions for alias-weighted precision.

The paper concludes that Tenure offers a coherent answer to the re-orientation tax through a schema that makes belief state auditable, a retrieval design that ensures precision, and a compaction architecture that maintains store integrity over time.