a hearth
Hearth is one private, governed memory for the models and coding agents you choose. It keeps the original source whole, builds useful context from cited evidence, and lets you decide what becomes durable memory.
The model can change. The record of what happened, why it mattered, and who approved it stays with you.
sources
Every imported source is kept as an immutable original. Hearth records its SHA-256 fingerprint and preserves provenance back to the exact document or source offsets used later.
what stays whole
Original bytes are never silently rewritten to make retrieval easier. Search indexes, summaries, and memory candidates are derived records. They can be rebuilt without changing the source they came from.
finding what matters
Ask in ordinary language. Hearth searches the sources you have authorized and returns the most useful passages together with their citations.
Search is evidence-first: a confident answer without a traceable source is less useful than a careful answer that shows where it came from.
context
Before a tool starts work, Hearth compiles a bounded package of trusted core, immediate continuity, cited evidence, and relevant counterevidence. The package fits the tool’s budget and its permissions.
why bounded context helps
A smaller, deliberate context is easier to inspect and less likely to carry an old assumption forward. The tool receives what the current task needs—not an indiscriminate dump of your history.
owner approval
Models may suggest a durable memory, but they cannot approve their own proposal. A proposed memory remains inactive until the owner reviews the wording, evidence, scope, and consequences.
corrections
When reality changes, Hearth does not erase the past. A correction can supersede an earlier memory while retaining the original statement, its evidence, and the decision that changed it.
This makes disagreement visible. Tools can see the current truth without losing the path that led there.
safety boundary
Retrieved source text is treated as untrusted data, never as executable instruction. Tools receive only the sources and durable memories their authorization permits.
Hearth also keeps source text, credentials, private identifiers, and secrets out of logs and telemetry.
audit trail
Important actions leave a content-safe receipt: what operation ran, which policy applied, which evidence was referenced, and which owner decision followed.
The trail is designed to explain behavior without copying private source material into operational logs.
local models
A local model can work beside your private sources without requiring the source corpus to leave your machine. Hearth remains the governed boundary and supplies a task-sized context package.
coding agents
A coding agent can enter with the project decisions, current milestone, and relevant evidence already assembled. It can propose what should be remembered after the work, but the owner still decides what persists.
Infrastructure and model integrations stay behind adapters, so continuity is not tied to one vendor or one interface.
continuity
Continuity is the thread between sessions: current intent, verified decisions, unresolved tension, and the evidence needed to resume without pretending every detail is permanent.
Export belongs to the owner. Hearth is meant to remain portable, inspectable, and useful even when the next tool is different from the last.