In one sentence:
Portable, provenance-first context that works across GPT, Claude, and local models, no lock-in.
Why it matters:
Your strategy survives tool & model changes; outputs stay comparable on the same history.
- Portable across GPT, Claude, and local/Ollama.
- Receipts on every recall (who/what/when/which model/source).
- Role- and project-threaded recall; not another notes app.
See it on the demo ·
Methods (Pilot-only)
What “agnostic” means in practice
- Same history produces comparable answers across models.
- Provenance (receipts) travels with the recall.
- No provider-specific tools are required to replay.
- Clear pass rules so teams can decide when a swap is “safe.”
Examples
- Switch from GPT → Claude without losing decision rationale.
- Work offline on a local model and keep the same memory thread.
- Handoff to a teammate: they see what changed, who decided, and why instantly.
FAQ
What does “LLM-agnostic” mean? Your memory and context work with any model, GPT, Claude, local/Ollama, and future models, without provider lock-in.
Why not just use one provider’s memory? Portability + provenance. Threads survive tool changes, and every recall carries who/what/when/which model so decisions remain auditable.
How is this different from a notes app? It’s a recall + provenance layer tied to roles and projects. It feeds the right slice back into whatever model you use next.
In one sentence: A compact export of the relevant thread (decisions, sources, role) you can replay on any model.
Why it matters: Enables apples-to-apples comparison (GPT ↔ Claude ↔ local) and quick handoffs.
See it on the demo
In one sentence: Memory that carries its own receipts: each recall points to source, who, when, and (if AI-assisted) which model.
Why it matters: Prevents contextless answers, makes model swaps auditable, and enables stable comparisons on the same history.
Methods (Pilot-only) · See MPR
In one sentence: A tiny JSON stamp attached to each recall showing who/what/when/which model/source (plus optional integrity hash).
Why it matters: Makes answers auditable and reproducible across LLMs.
Under the hood (gated)
In one sentence: Mix lexical/regex with vectors and filters to cast a smart net before assembly.
Why it matters: Reduces junk; surfaces the right context with fewer tokens.
See it on the demo
In one sentence: The “receipts” carried with each recall, who decided, what changed, when, and which model or source was used.
Why it matters: Makes AI-influenced work auditable and speeds up reviews and handoffs.
See also: Provenance-first · Provenance chip (MPR)
In one sentence: Order candidates after retrieval to promote, dedupe, and diversify before capping.
Why it matters: Tight, relevant context > maximal context; improves answer stability.
Signals we consider include relevance, recency, source quality, and role fit; specifics are covered in the Pilot.
In one sentence: Memory stays stable as you change models; you can compare outputs on the same slice.
Why it matters: Prices, policies, and capabilities shift, your process shouldn’t.
See it on the demo
Comparable if… (checklist)
- Same memory slice (unchanged).
- Same universal prompt style and role/objective (if used).
- Deliberation budget capped to avoid runaway chains.
- Pass rules met (e.g., preserves key timeline events; next step aligns).
In one sentence: Pull the “why, who, and what-changed” in under 60 seconds for the current role and project.
Why it matters: Kills re-explaining loops; turns memory into a working surface, not an archive.
Typical pattern
- Filter by project and role to surface the last relevant events.
- Summarize facts → rationale into a quick timeline with receipts.
- Propose the next step consistent with the stated objective.
In one sentence: Decisions are stored with links to sources and owners, not just summaries.
Why it matters: Explains “who approved this?” without meetings.
See receipts
In one sentence: Keep Founder/PM/Researcher threads clean so each recall uses the right perspective, without brittle prompts.
Why it matters: Reduces noise; improves relevance without prompt gymnastics.
Implementation notes
- Tag events with role at capture time.
- Switch roles via UI/metadata rather than prompt text.
- Merge conflicting roles only on explicit request.
In one sentence: Capture happens passively as you work, decisions and sources flow into the memory thread.
Why it matters: No new workflow; less copy-paste; fewer gaps.
See it in the Approach
In one sentence: Run the same slice on different models to compare outputs or reproduce a result later.
Why it matters: Transparency for stakeholders; confidence to change vendors.
See it on the demo
Comparison rubric (quick)
- Preserves at least two key timeline events (what, who, when, and why).
- Proposed next step aligns with the stated objective.
- Receipts are present; no fabricated sources.
- Minor wording differences are acceptable.
In one sentence: A timeline of AI-influenced changes with owners, sources, and timestamps.
Why it matters: Answers “who approved this?” without meetings; keeps scope honest.
We log timestamp, change description, rationale, owner, source URI, and (if relevant) model and status.
In one sentence: General phrase people use for saved context in an AI tool, often locked to one provider.
Why it matters: If it’s provider-bound, you still lose history when you switch models. See “LLM-agnostic memory.”
See it on the demo
In one sentence: The amount of text a model can consider at once (token limit).
Why it matters: Bigger windows help, but they’re not memory; portable recall outlives any single chat. See also context rot.
See it on the methods · See it in the Approach
In one sentence: Degrading answer quality as bloated context windows accumulate noise and stale detail.
Why it matters: We cap and compact to keep recall tight and comparable across models.
See it on the methods
In one sentence: Fetch documents and feed them to a model at prompt time.
Why it matters: Useful for facts, but you still need provenance and a cross-model memory for decisions and timelines.
See it on the demo
In one sentence: A hard cap on “thinking” to prevent overlong chains and drift; lock early once quality is met.
Why it matters: Keeps runs reproducible; controls cost and variance across LLMs.
We cap steps/tokens and lock early when pass rules are satisfied; specifics are covered in the Pilot.
In one sentence: A personal knowledge system for notes and ideas.
Why it matters: Great for writing; different from replayable, cross-model project memory with receipts.
See it in the Approach
In one sentence: Structured why / who / source / date attached to a change so outputs are auditable.
See also: Provenance · Provenance chip (MPR)
In one sentence: Replay passes when ≥2 key timeline events are preserved and the proposed next_action is aligned; phrasing differences are OK.
See also: Model-swap safe · Replay
In one sentence: A tool-calling standard some chat apps use; optional and region/admin-dependent. Threadbaire works without it.
Approach: Stack fit (integrations)
