← Clarence case study Public-safe architecture · July 15, 2026
Clarence system report

How Clarence remembers

Memory is not one database. It is a governed stack of prompt context, transcript recall, structured knowledge, semantic indexes, a human-readable Wiki, procedural skills, and temporary execution state.

The system is live and useful. Its value comes from separating those jobs instead of calling every stored thing memory. Clarence decides what becomes durable, where it belongs, and what must remain temporary.
1,238Current memories
10,259Current facts
1,863Indexed vault notes
4Local vector lanes

Verified July 15, 2026. Counts are a reviewed snapshot, not real-time telemetry.

01 · Architecture

The current system

A message enters through Hermes. Bounded context is assembled for the turn. Durable knowledge stays in separate stores with separate write rules. Read-only clients do not become peer memory writers.

Solid green lines write · dashed orange lines read · dotted gray lines stage or report

Public-safe Clarence memory system architecture A layered diagram of the Hermes runtime, prompt context, session history, retrieval tools, structured memory, Wiki, skills, run-state, corpus staging, maintenance, and read-only external clients. INTERACTION + RUNTIMETURN CONTEXT + HISTORYRETRIEVAL + STRUCTURED MEMORYREADABLE + PROCEDURAL + STAGINGMAINTENANCE + EXTERNAL READERS James + local surfaceschat · terminal · dashboard Hermes / Clarenceorchestration · judgment · canonical write gate External clientsbounded read or handoff lanes Prompt memorystable preferences · working rules Memory providerprefetch · mirror · stage Session historytranscripts · search · continuity History searchprior sessions Memory toolsexact · semantic · deep recallgoverned writes + retirement Context graphentities · facts · relationsread-only prototype Structured knowledge systemmemories · facts · entities · relations · vault index · corpusSQLite + four local semantic vector lanes System Wikihuman-readable compiled truth Skillsrepeatable procedures Run-statetemporary task continuity Corpus stagingmanual promotion path Sync + index + vectors Report-only hygiene audit Growth guard Read-only gateways

The structured database is authoritative for machine-queryable knowledge. The Wiki is authoritative for readable synthesis. Neither transcript history nor run-state becomes truth merely because it exists.

02 · Turn flow

How one conversation turn works

01
Hermes receives the message.

Identity, permissions, model policy, and the current profile establish the operating boundary.

02
Stable prompt memory is already present.

A small reviewed set of preferences and durable operating facts enters the session context.

03
The memory provider adds bounded recall.

Relevant structured memories and facts are retrieved for the turn without dumping the full database into the prompt.

04
The transcript enters session history.

Conversation history remains searchable for what was said, including goals, decisions, and prior corrections.

05
Additional tools retrieve evidence.

Exact lookup, semantic search, deep recall, Wiki search, and relationship queries serve different evidence needs.

06
The turn does not automatically become truth.

Most conversation remains transcript history. Durable promotion requires a deliberate destination and a governed write.

07
Explicit preferences can be staged.

Candidate capture identifies statements worth review, but candidates do not promote themselves.

08
Durable knowledge lands in the right layer.

Stable facts, readable synthesis, repeatable procedures, and temporary task state remain separate.

03 · Inventory

What each component does

The same information can appear in more than one layer, but each layer has one job. That is the difference between redundancy and confusion.

ComponentRoleReadsWritesStatus
Prompt memorySmall always-needed preferences, corrections, and stable system facts.Session-start snapshotGoverned memory toolactive
Memory providerAutomatic bounded prefetch, mirrored hot-memory changes, and candidate staging.Structured knowledgeMirrors and candidatesactive
Candidate stagingHolds possible preferences or directives until reviewed.Explicit user signalsCandidate rows onlystaging
Session historyRecords conversations, usage, routing, and continuity.Hermes runtimeEvery conversationactive
History searchFinds actual prior discussions without pretending transcript text is current truth.Session historyNoneaccess
Structured knowledgeCanonical machine-queryable facts, memories, entities, relations, indexes, and lifecycle state.Memory tools and maintenanceClarence-governed pathscanonical
Memory toolsExact lookup, semantic search, deep recall, governed writes, retirement, and feedback.Structured knowledgeApproved operationsactive
Vector lanesLocal semantic retrieval across memories, facts, vault notes, and large documents.Embedded source rowsIndexer-managed vectorsfull coverage
Deep recallCombines semantic, lexical, metadata, Wiki, profile, and corpus candidates.Multiple retrieval lanesTelemetry and feedbackactive
Context graphAnswers relationship-first questions over explicit entities and relations.Structured knowledgeNoneread-only pilot
System WikiHuman-readable compiled truth: architecture, decisions, source maps, and current state.Task intake and reviewDurable synthesiscompiled truth
Vault indexMakes reviewed notes searchable beside structured memory without pretending they are identical.Reviewed notesSearch index and vectorsscheduled
SkillsProcedural memory for repeatable workflows, verification, and tool use.Skill catalogReviewed skill filesprocedure authority
Run-stateTemporary task handoff, command receipts, blockers, and next action.Current task onlyDuring substantial worktemporary
Corpus indexLarge-document metadata, chunks, exclusions, extraction state, and semantic vectors.Curated source packsManual promotion pathactive store
Maintenance jobsRefresh indexes, audit hygiene, checkpoint growth, and report drift.Stores and schedulersIndexes or reportsscheduled
External readersGive approved clients a narrow view without granting canonical write authority.Public-safe or read-only surfacesNonebounded
04 · Authority

What wins when layers disagree

Source hierarchy

Use the narrowest authoritative layer

  • Live services and source code establish runtime truth.
  • Prompt memory holds stable preferences and corrections.
  • Structured knowledge holds queryable durable state.
  • The Wiki holds readable synthesis and design reasoning.
  • Skills hold procedures.
  • Session history proves what was said.
  • Run-state proves what happened during a task.
Write hierarchy

Clarence remains the gate

  • External clients read or submit handoffs.
  • Candidate staging never self-promotes.
  • The graph has no write path.
  • Hygiene audits propose rather than silently mutate.
  • Retrieved material is evidence until verified.
  • Public surfaces receive curated snapshots, not private live memory.
05 · Retrieval

Six ways to find the right evidence

Lane 01

Exact lookup

Known names and identifiers avoid semantic guesswork.

Lane 02

Semantic search

Local vectors retrieve related memories, facts, notes, and document chunks.

Lane 03

Lexical backstop

Full-text, title, path, and metadata affinity recover exact evidence vectors miss.

Lane 04

Deep recall

A bounded candidate pool combines lanes and protects strong exact hits.

Lane 05

Relationship recall

The graph answers dependency, ownership, and project-membership questions.

Lane 06

Feedback

Useful and noise ratings create evidence for improving ranking over time.

06 · Status

Active, experimental, and retired

Active core

What the system depends on

  • Hermes orchestration
  • Prompt memory and session history
  • Structured knowledge and local vectors
  • System Wiki and skills
  • Scheduled indexing and reporting
Active support

What makes it usable

  • Automatic bounded prefetch
  • History search and deep recall
  • Vault and corpus indexes
  • Hygiene audit and growth monitoring
  • Public-safe graph and Pulse surfaces
Experimental or staging

Useful without being authority

  • Explicit-signal candidate capture
  • Context graph pilot
  • Manual corpus staging
  • Optional reranking path
  • Early retrieval-feedback loop
Retired

Do not revive by accident

  • The former OpenClaw runtime
  • The old default-profile control plane
  • Automatic conversation distillation
  • The former remote memory replica
  • Historical cron piles and hidden routing assumptions
07 · Open work

What still needs discipline

Transcript retention

Conversation history grows quickly. Indexing and checkpoints are working, but long-term retention still needs an explicit policy.

Prompt scarcity

Always-injected context is finite. New facts must replace or consolidate older material instead of accumulating forever.

Review debt

Hygiene is intentionally proposal-first. The remaining work is reviewing and applying the right proposals without turning maintenance into autonomous deletion.

Ranking maturity

Semantic coverage is complete, but optional reranking remains disabled and the feedback corpus is still young.

Corpus freshness

Large-document promotion is manual. That keeps publication deliberate, but it also creates a second freshness clock.

Legacy naming

Parts of the durable storage layer still carry names from the previous runtime. The system works, but the namespace can mislead future audits.

08 · Public boundary

What this report leaves out

This is a curated architecture report. It excludes private memory text, personal profile content, conversation excerpts, work-item titles, filesystem locations, network topology, credentials, operational thresholds, precise activity times, and writable remote interfaces. The point is to show the design without publishing the operating record.