Agents shouldn’t have to reconstruct current state from raw history every time they plan.
Belief Stack maintains state between evidence and planning — producing compact projections for agents and inspectable traces for humans.
One substrate. Two surfaces.
Agents get the minimum state needed to act. Humans get the trace needed to inspect why.
What the system believes, why that belief is warranted, whether it still holds, and what changed over time.
Belief Stack tracks working assumptions over time: what a system treats as true, why that belief is warranted, whether it still holds, and what changed.
For models, that state becomes a compact overlay.
For humans, it becomes an inspectable trace.
Logs show what happened.
Belief state shows what the system was relying on.
Maintained state is a planning primitive.
Belief Stack is the substrate layer between evidence and planning. Tested directly in v0.3 →
Four consequences of maintained state. Two measured by the experiments documented below; one structural to the architecture; one whose net axis is still open.
Fewer input tokens per planner call. Net runtime economics still depend on extraction and maintenance costs.
~10× smaller input · net cost: v0.4b open
Faster planning per call. Users experience this directly.
~3.2× faster (v0.3) · ~1.4× cross-model (v0.4c1)
Preserves what the agent relied on, why it was warranted, and when it changed.
human audit surface · governance: untested
Improves planning correctness on tested operational workflows.
+9.7 pp average cross-model lift
The Belief Stack
The substrate layer between evidence and planning. One substrate, asymmetric consumer projections — sparse for planners, rich for humans. Empirical-status section now carries the v0.2.2, v0.3, v0.4a, and v0.4c1 results.
Belief Stack v0.4c1 — cross-model replication
Across four LLMs from three providers, sparse maintained-state projections reached 99.0% planning correctness at 241 mean input tokens, vs 89.3% at 2,502 for raw history. The thesis held on every model tested; compression-vs-substrate isolation narrowed to model-dependent. The third pre-registered challenge the thesis survived.
Belief Stack v0.3 — planning-side experiment
A 285-token belief overlay (14% of the raw-log baseline) reached 98.7% planning correctness vs. 90.7% for the strong-baseline raw-context arm, at 3.2× lower latency. The smallest arm won. The model was overburdened by reconstruction, not under-informed by less context.
The Recent Log Is Not the Same as State
What goes wrong when a long-running assistant has to answer questions about workflow state from raw log context alone — and what changes when a maintained belief overlay is added.
Watching an Assistant Forget
TKOS Log-Replay v1 — preregistered offline audit of 164 Claude session logs, 20,190 evaluation turns, with TP/FP/FN/TN accounting. The empirical substrate behind Operational Belief-State Grounding v0.1, v0.2.2, and v0.3.
Reading the TopicSpace Belief Field
Sensemaking v1.5 — the Belief Stack pattern applied to financial-market narrative pressure across 31 actors and 173 days. Near-chance aggregate calibration with measurable regional heterogeneity.
TopicSpace began as a market sensemaking system tracking AI-ecosystem narratives across public companies. That work now remains online as an archived research surface and as the empirical substrate for the Sensemaking v1.5 case study. Daily market commentary is no longer updated; historical pages remain available with archive banners.