Conversation is temporary
Important decisions disappear into old threads unless they are captured as durable, navigable knowledge.
Local-first AI continuity
Your AI starts where you left off.
AutoCache Continuity turns user-owned knowledge into durable context across models, agents, and sessions. It combines the Autonomous engine, structured Volumes, and receipt-backed upkeep without turning your memory into a black box.
Continuity for working knowledge, not prompt-response caching. No personal Volumes, credentials, or live operational data are bundled.
$ resume work
01
Navigate current knowledgecatalog / topic / exact range / relations
found
02
Restore active contextgoal / ownership / verified progress / next action
loaded
03
Check knowledge integritynavigation / sidecars / freshness / receipts
green
Context restored
Continue from evidence, not chat-history guesswork.
Proof is staged
- 1Source found
- 2Installed
- 3Discoverable
- 4Invoked
- 5Behavior proven
The continuity gap
A longer context window is not a memory system.
Chats end. Models change. Files drift. Search finds words without knowing which record is current. AutoCache adds a governed continuity layer between your knowledge and the AI using it.
Retrieval needs authority
Finding a similar passage is not enough. Current records, supersession, provenance, and exact scope matter.
Automation needs evidence
A green scheduler status does not prove useful work. Every unattended stage needs its own output and receipt.
The system
One continuity layer. Four explicit parts.
Events provide timing, skills provide procedure, tools provide capability, and Volumes hold durable truth. The parts cooperate without pretending that one layer proves another.
Inputs
Models, agents, files
Work enters from supported AI surfaces and reviewed sources.
Engine
Autonomous
Recall, extraction, continuity, ownership, and evidence-aware operations.
Durable layer
Volumes
User-owned knowledge with catalogs, section maps, relations, and provenance.
Autonomous
Gives supported AI clients a consistent way to navigate knowledge, recover session state, track multi-step work, capture verified learning, and check propagation after a change.
Volumes
Portable, readable files organized by topic. The knowledge remains inspectable outside any single model, vendor, database, or chat interface.
Skills
Teach each AI how to recall, verify, write, coordinate, and hand off work. Skills travel with the method instead of relying on a model to remember the rules.
Hooks
Route procedures at lifecycle events when the host supports them. Installation is not enforcement; trust and a real route-complete event must be tested separately.
Volumes
Your knowledge stays readable, portable, and yours.
Volumes are the durable substrate beneath AutoCache: markdown-first knowledge organized into topics, with machine-readable navigation and relations layered beside the content.
No private corpus in the product package.
Starter material is synthetic. Your real knowledge is added separately and remains under your control.
Catalogs route the topic
A compact index points the AI to the canonical subject instead of making every session scan everything.
Operating files hold current truth
Topic records consolidate decisions, procedures, discoveries, status, and their evidence over time.
NAV maps keep reads bounded
Section ranges let an AI open the relevant passage directly, reducing noise and avoiding full-corpus reads.
Sidecars connect the knowledge
Separate metadata companions carry keywords, provenance, supersession, and relationships without rewriting the source prose.
Search follows structure
Literal, semantic, and concept search reinforce navigation when needed; they do not silently outrank canonical records.
The method
Recall first. Learn only when it earns permanence.
AutoCache is not an indiscriminate transcript dump. It separates temporary context from durable knowledge and uses verification before promotion.
Navigate
Enter through the catalog, then the canonical topic, exact section, and relation layer.
Act
Restore goal, ownership, verified progress, risks, and the exact next move.
Verify
Read back writes, parse structured artifacts, and test behavior at the layer being claimed.
Capture
Keep a learning only when it is reusable, specific, and genuinely new.
Propagate
Update navigation, relations, and dependent references so one change does not fork the truth.
Scheduled upkeep
The memory engine includes an integrity loop.
Knowledge systems decay when new material lands but navigation, relations, and health signals do not follow. AutoCache packages that upkeep as previewable, separately activated work.
AutoCache
fresh context
receipt-backed
1Detectchanged or stale
2RefreshNAV and relations
3Reconcileperiodic full pass
4Proveoutput and receipt
Change funnel
File events identify what changed so routine refreshes can stay bounded instead of rescanning the whole corpus every time.
Navigation and relation refresh
Changed knowledge gets updated section maps and sidecar metadata, keeping the route to current truth intact.
Self-heal with a full backstop
Incremental work retries unresolved drift; periodic full reconciliation catches anything a change detector missed.
Terminal receipts
Each unattended run records its own result and provenance. Health consumers read the current receipt contract before falling back to legacy logs.
Memory integrity pack
Change detection, navigation refresh, sidecar upkeep, drift checks, self-heal, and receipt wiring belong with the memory engine.
- Preview before registration
- Ship disabled by default
- Validate host prerequisites
- Activate in a separate step
Reasoning pack
Deeper audits, summaries, and synthesis can use a supported model subscription, but that dependency is never imposed silently.
- Explicit model dependency
- Separate setup and cost boundary
- Independent outputs and receipts
- Safe to omit from the base install
Personal and domain automations
Trading, backups, business workflows, and machine-specific jobs keep their own identity and do not become AutoCache merely because they use its knowledge.
- No silent product bundling
- No cross-domain ownership takeover
- No machine paths baked into the core
- Connect through explicit interfaces
Trust boundaries
Capability is visible. Authority stays explicit.
AutoCache treats source, installation, discovery, invocation, and enforcement as different states. It reports what is proven without turning configuration into a security claim.
Local-first by design
Private knowledge is not bundled into the product. The receiving user decides what enters their Volumes and which clients may reach it.
Human gates remain
Restarts, irreversible actions, credentials, money, and other high-impact decisions remain explicitly authorized.
Hooks do not self-certify
A hook becomes meaningful only after its exact definition is trusted and a harmless real event proves the route that matters.
Schedules do not self-certify
Task status, process launch, and timestamps are intermediate evidence. Semantic output and a run-specific receipt close the proof chain.
Readable before proprietary
Markdown and explicit metadata keep the knowledge inspectable. Optional indexes accelerate recall without becoming the only copy of truth.
Current truth outranks history
Superseded material is marked, dated evidence is reconciled, and mutable facts are rechecked at the live source.
Private preview
Build continuity you can inspect.
AutoCache Continuity is being developed privately while the runtime, plugins, Starter Volumes, and maintenance pack are validated as separate, evidence-backed stages.