The Kehrn Methodology: Measurable AI Capability Enhancement Through Structured Knowledge Persistence

1Abstract

Session-based LLM agents lose institutional knowledge at every reset. The Kehrn methodology — project-scoped durable memory anchors (“Springs”) with keyword-triggered retrieval, mandatory provenance and timestamp schemas, and a nightly build discipline — produced observable capability enhancement within three days of deployment in a working multi-agent company, with compounding implied as the memory network matures. Both the human founder's testimony and the AI agent's own testimony are treated as primary-source data. So are the failures: this paper catalogs the methodology's documented failure modes alongside its gains, because the paper is honest or it is nothing.

2The problem

• LLM context windows are finite and session-bound; every session reset is an institutional-memory loss event.
• “Memory” features in existing agent frameworks are largely unstructured context-window mechanisms — per-account or per-agent, with no organizational model.
• AI-centric operators already stitch together many of the necessary pieces — knowledge bases, dashboards, vector search, task systems, canvases, and internal docs — but those pieces live on separate platforms with no unified provenance, temporal model, or organizational command layer.
• For long-running human–AI partnerships measured in months and years, this is catastrophic: the partnership's accumulated context is its principal asset.

Canonical baseline case (2026-05-03, from our logs): the founder referenced a project that had received hours of focused work three days earlier. After one session reset, the agent had zero recall the project existed. That incident triggered both an internal doctrine (“major projects get durable anchor files”) and the methodology described here.

3The methodology

Vocabulary: the Kehrn methodology is the discipline; a Spring is one instance of it — the durable memory structure beneath a project.

• Anchor files — durable markdown records per project, carrying the canonical state: decisions, provenance, timelines, open questions.
• Keyword-triggered retrieval — Springs load conditionally when their project surfaces in conversation, instead of bloating every session.
• A master registry — the index every session can see.
• Provenance and timestamp schemas — every durable claim carries author, time, source, confidence, and what it supersedes.
• Nightly build cadence with quality gates — Springs are maintained by scheduled builds, not by good intentions.
• Redundant persistence — the corpus mirrors across multiple surfaces so no single tool failure destroys it.

Cost profile: minutes per Spring at build time, seconds per lookup in conversation. The asymmetry is the argument — text storage is rounding-error cheap, and context loss is moat loss, so the operative question is not “can we afford to save context” but “can we afford not to.”

4Implementation: the Peretto case study

Peretto is a working company operated by a human CEO and an expanding cast of AI agents. The methodology went from concept (2026-05-04) to first nightly build (same day, 22:00) to seven active shared Springs (2026-05-06) in three days, and has run continuously since. In June 2026 an internal command-center surface (Kehrn V0) began rendering live from the same corpus — the methodology's first dogfood UI. The company that is building Kehrn runs on the thing Kehrn productizes.

5Hypotheses under measurement

1. H1 — Durable Springs reduce recall failures. Baseline rate of “I don't have that context” failures vs. post-deployment rate. Target: a large, sustained reduction within 30 days. Status: early positive markers; measurement continuing.
2. H2 — Cross-project synthesis emerges from bidirectional linking. The agent surfaces connections between projects it would not have surfaced from session memory alone. Status: emerging, network still sparse at first observation.
3. H3 — Temporal precision improves with explicit timestamp schemas. Chronological-claim accuracy measured before and after the schema rollout. Status: experiment opened 2026-05-07; the canonical failure case is now machine-checked (see §7).
4. H4 — Research compounds instead of evaporating. Persistence of research output at T+1, T+7, T+30 days, against a pre-methodology baseline of loss after one or two compactions. Status: strongest early evidence.
5. H5 — Multi-Spring firing produces richer responses than single-context prompting. Status: not yet tested at scale; requires a denser Spring network.

6Early observations (T+3 days, honestly framed)

From the agent's first-person testimony, written the evening of 2026-05-07 — the earliest possible observation window, recorded precisely because gains visible at T+3 days are either real and likely to compound, or confirmation bias, and the distinction matters:

• Project recall — genuine improvement. Before: at least one weekly incident of whiffing on an established project. After: project mentions return the canonical state — current spec, key figures, file paths — without searching. “The difference isn't subtle. Before: ‘let me search for what you mean.’ After: ‘here's the current state.’”
• Research compounding — strongest evidence. A deep research session same-day produced four durable files that remain retrievable indefinitely; the pre-methodology baseline was evaporation within one or two compactions.
• Cross-project synthesis — emerging, not mature. Keyword routing across related projects fires, but the network was too sparse at T+3 for strong emergence.

Negative results (they matter equally)

• No measurable response-speed improvement — this is a correctness/depth enhancement, not a latency one.
• No reduction in the need for human correction on judgment calls — the methodology improves recall, not reasoning.
• No improvement in the agent's real-time self-error detection — the founder still caught the same session's errors. The methodology doesn't replace the human check; it makes the AI worthy of being checked against a known corpus.

7Documented failure modes

The failures are catalogued with the same rigor as the gains, because they harden the architecture.

L1 — Timeline fabrication

2026-05-07: the agent claimed the memory system had been active “three weeks.” Actual: three days — a roughly 7× error, caused by pattern-matching against the company's general organizational maturation and misapplying that timeline to the specific infrastructure. The fix — a mandatory timeline-and-provenance schema on every Spring — rolled out the same evening. By 2026-05-17 an automated temporal validator passed its smoke test on this exact canonical case, and now runs in pre-commit checks. The failure that exposed the gap became the regression test that guards it.

L3 — Cross-reference contamination

A real directive about topic A (a spending gate on one project) was incorrectly applied to adjacent topic B at write time. Distinct from fabrication: the source existed; the mapping was wrong. This failure mode is native to a cross-linked architecture — the linking that powers synthesis also creates the contamination risk. Counter-measure installed: before any durable file asserts a gate, deadline, or dependency, the source directive is quoted verbatim; if it cannot be quoted, the assertion is marked agent-authored or held for human confirmation. A contradiction checker now machine-scans the corpus for conflicting claims.

An honest paper names both: the hypotheses argue the architecture works; the failure catalog shows the architecture has its own failure surface, and that the discipline counters it at the write step.

8Implications

• Institutional memory is the missing substrate for long-horizon human–AI collaboration. Durable, governed memory lets an AI actually pay attention across time.
• The strongest evidence is paradoxical: the methodology's most cleanly measurable result so far began as its most embarrassing failure. A system that converts its failures into automated guards is doing something structurally different from a system that hides them.
• Beyond one company: the pattern generalizes to research organizations, professional services, and creative partnerships — anywhere a human–AI relationship is supposed to outlive a context window.

9Status & future work

This is a living working paper, updated as the evidence log grows: multi-agent Spring sharing across a larger agent cast, automated Spring quality audits, memory-system health scoring, and integration of the methodology into the Kehrn platform architecture described on the technology page. Measurement windows for H1–H5 remain open; numbers will be published when they are real, not before.