Context LatticeBy Private Memory Corp
Operator wiki

The complete public/main operating manual for ContextLattice.

One place for launch commands, endpoint behavior, retrieval mode policy, continuation mechanics, and production-safe workflows.

Designed for humans and agents running real systems.

Start in 5 minutes Endpoint atlas Retrieval modes Playbooks
Quick path

Fastest route from install to reliable memory retrieval

  1. Launch with gmake quickstart and choose lite or full profile based on machine resources.
  2. Validate /health and authenticated /status.
  3. Run one /memory/write + one /memory/search smoke call.
  4. Use fast mode first; escalate to balanced or deep only when scope or recall quality requires it.
  5. On partial results, use continuation token/events instead of re-running expensive full-depth queries immediately.
Visual atlas

How best-tools routing works in production

ContextLattice wiki tool atlas
Tool Atlas: orchestrator ingress, fast-now sources, slow async continuation, and evidence return path.
Ingress

Single API surface

  • Primary: http://127.0.0.1:8075
  • Auth: API key required for protected routes
  • Runtime: Go orchestrator + Rust retrieval hot paths
Fast-now plane

Low-latency answers

  • topic_rollups
  • qdrant + weaviate
  • postgres_pgvector
Continuation plane

Depth without blocking UX

  • letta + memory_bank + mongo_raw
  • Continuation token + SSE events for async completion
  • Fail-open policy preserves immediate responses
Endpoint atlas

Best tools to call, by intent

Intent Endpoint When to use Output shape
Write checkpoint POST /memory/write Persist a decision/result during long tasks ok, sink fanout status, queue details
Context retrieval POST /memory/search Scoped recall before planning/coding/inference Results + grounding + retrieval debug + continuation metadata
Hierarchical context pack POST /memory/context-pack Multi-file/topic synthesis for complex tasks Structured pack with topical grouping
Agent session rollup GET /v1/agents/sessions/{id}/rollup Summarize current objective state, evidence coverage, risks, and next action Contracted agent_session_rollup.v1 payload with runtime confidence
Next-model context package GET /v1/agents/sessions/{id}/context-package Prepare a bounded reference packet before a hard model handoff or follow-up prompt Contracted agent_prompt_context_package.v1 payload plus reference_prompt
Capability discovery GET|POST /tools/skills_index_search Find relevant skills/tools without loading every capability into startup context Ranked Skills Index matches with paths, score, and matched terms
Typed memory edges GET|POST /v1/memory/edges, POST /v1/memory/edges/backfill Persist explicit relationships and audit/apply deterministic retroactive or opt-in inferred edges Durable edge records in the local memory graph index with dry-run counts and bounded inferred scoring
Graph neighbors POST /v1/memory/neighbors Detail dives from a known memory id Explicit edge neighbors merged with semantic/topic neighbors
Health GET /health, GET /status Preflight and operator checks Service health, queue/retention/runtime ownership
Async continuation GET /memory/search/continuations/{token}/events Stream slow-source completion events after staged fetch SSE lifecycle updates + completed source payloads
Retrieval policy

Mode selection: speed first, depth on demand

Fast mode

  • Use for: iterative engineering loops and narrow scoped reads.
  • Behavior: prioritize fast-now plane, defer deep sources.
  • Timeout guidance: ~25s client budget.

Balanced mode

  • Use for: most default operator/agent reads.
  • Behavior: mixed fast-now + measured continuation.
  • Timeout guidance: ~60s client budget.

Deep mode

  • Use for: long-horizon recall, audit, migration, root-cause.
  • Behavior: maximizes recall breadth; continuation still applies.
  • Timeout guidance: ~75s client budget.

Mode ladder

flowchart LR A[Need context] --> B{Scope known?} B -- yes --> C[Fast mode with project/topic] B -- no --> D[Balanced mode broad project] C --> E{Enough coverage?} D --> E E -- yes --> F[Proceed + checkpoint writes] E -- no --> G[Deep mode + continuation stream] G --> H[Use neighbors/context-pack for detail dives] H --> F
Playbooks

Production runbooks by job-to-be-done

PB-01 · Launch

Bring up safely and verify in under five minutes

  • gmake quickstart chooses profile + launches services.
  • Run /health then authenticated /status.
  • Smoke test one write and one search before real traffic.
PB-02 · Retrieval quality

Increase recall without overpaying latency

  • Start scoped + fast, widen only when coverage is insufficient.
  • Use continuation stream instead of blocking every request on slow lanes.
  • Persist concise checkpoints so future queries hit high-signal facts first.
PB-03 · Incident response

Fail-open by design, recover without losing context

  • Use /status and dashboard to identify degraded sources.
  • Continue operations on healthy fast-now sources while continuation queues warm slow lanes.
  • Escalate only when protected/non-degradable lanes are unhealthy.
PB-04 · Release

Promote only with evidence

  • Run benchmark + recall quality + soak checks.
  • Tag release notes with behavior changes and operator guidance.
  • Keep deterministic artifacts and changelog links in each release.
Agent integration

Copy-ready templates for Codex, Claude, OpenCode, Hermes

Use these templates to get consistent ContextLattice usage behavior across popular agent interfaces:

  • docs/public_overview/templates/AGENTS.contextlattice.md
  • docs/public_overview/templates/SKILLS.contextlattice.md
  • docs/public_overview/templates/agents/codex.md
  • docs/public_overview/templates/agents/claude-code.md
  • docs/public_overview/templates/agents/opencode.md
  • docs/public_overview/templates/agents/hermes-agent.md
Release readiness

Public/main checklist before each post or launch

  1. All critical paths healthy on /status with no unresolved alerts.
  2. Write/read smoke test passes with API-key auth.
  3. Continuation metadata present for deferred slow sources.
  4. README links, website nav, and release notes updated together.
  5. Tagged release exists and matches the intended commit SHA.