Episodic memory in PKM is the deliberate capture and structuring of judgment — why you made decisions, how they turned out, what failed and why — as a distinct vault layer from the facts that most PKM systems store. The concept is drawn from Endel Tulving's (1972) psychological distinction between semantic memory (facts, concepts) and episodic memory (events, contexts, personal experiences). Koylan's Personal OS (2026) is the first practitioner architecture to make the split first-class: alongside content/contacts/metrics files (semantic), the system maintains experiences.jsonl, decisions.jsonl, and failures.jsonl (episodic).
The architectural distinction: "There's a difference between an AI that has your files and an AI that has your judgment."
The Three Log Types
Koylan's concrete schema:
experiences.jsonl
- Key moments in your life/work
- Emotional weight score (1-10) — intensity of the experience's impact on you
- Used to surface patterns across life — what kinds of experiences recur, which ones were pivotal
decisions.jsonl
- Key decisions, not just what you decided but how you decided
- Fields: options considered, reasoning for each, constraints, outcome (tracked retroactively)
- Example from Koylan: the Antler Canada $250K investment vs Sully.ai Context Engineer role decision — both options, reasoning, applied priorities ("Learning > Impact > Revenue > Growth"), and the framework ("Can I touch everything? Will I learn at the edge? Do I respect the founders?")
failures.jsonl
- What went wrong; root cause; prevention steps
- Koylan: "the most valuable [log]; encodes pattern recognition that took real pain to acquire"
Why It Matters
Semantic PKM is incomplete
A vault full of highlights, atomic notes, and reference material tells an agent what the user knows. It does not tell the agent how the user thinks. Two people with identical reading material make different decisions; the difference lives in episodic memory, not semantic.
Decisions are the high-leverage data
Most of what a personal AI agent should do is make or support decisions — which task to prioritize, whom to contact, whether to accept an opportunity, how to respond to a message. Decisions draw from experience, not facts. A vault that logs decisions is a vault that an agent can reason from.
Failures are under-captured
Most people never systematically log their failures. The failures that matter are the ones with a story attached (what happened, why, what they'd do differently), which is usually told once and then lost. Explicit failures.jsonl makes this a first-class vault layer.
Emotional weight gives retrieval signal
Koylan's 1-10 weight on experiences is a salience signal agents can use to surface what matters. Without it, all experiences are equally weighted and the agent cannot prioritize.
Relationship to Existing PKM Concepts
- Daily notes — often contain episodic material but unstructured; episodic memory in the Koylan sense is distilled and tagged rather than chronological narrative
- Evergreen notes — semantic, not episodic; the distinction complements rather than replaces
- Personal Identity — identity notes reference patterns in experience/decision logs; they are the synthesis layer on top of raw episodic material
- Dots — the OSK dots framework with its wisdom, hypothesis, and similar types overlaps with the episodic-memory concept, often serving the same function
Implementation Patterns
Append-only
The logs are append-only (JSONL format). Decisions don't get deleted when they change; new entries supersede. Failures don't get edited; new insights become new entries referencing the old. This matters because:
- Pattern analysis benefits from full history
- Agent reasoning can trace how your thinking evolved
- You avoid the bug where an agent overwrites prior judgment
Retroactive outcome tracking
A decision is logged when made; outcomes are logged later as they unfold. This is a form of lightweight periodic review — asking "how did the decisions I made three months ago actually play out?"
Emotional weight is cheap
A 1-10 score per experience takes two seconds. It is dramatically cheaper than writing a paragraph of analysis and nearly as useful for agent retrieval.
Cross-reference to decision frameworks
If you have an explicit priority order ("Learning > Impact > Revenue > Growth" in Koylan's case) or decision frameworks, the decision log should reference them. Then an agent can apply your frameworks to new choices by pattern matching.
Risks and Caveats
- Privacy surface. Episodic logs are more sensitive than most other vault content — they reveal how you think, what you fear, what you regret. Access control and backup discipline matter more.
- Not for everyone. Highly structured decision logging is effortful. Practitioners who try it briefly often stop. The value compounds, but only if the habit survives.
- Confirmation bias. Logging decisions with reasoning can ossify wrong frameworks if they're never challenged. Periodic review should include asking whether the logged reasoning was actually correct.
- Retroactive reconstruction is degraded. Logs written after the fact from memory are lower-quality than logs written at the time of decision. This is the PKM version of survivorship bias.
Practical Starting Point
If you want to try this pattern without a full Koylan-style architecture:
- Create three notes (or three JSONL files):
experiences,decisions,failures - Every ~2 weeks, add 1-3 entries to each
- For experiences, include a 1-10 weight
- For decisions, include options considered and the framework you applied
- For failures, include root cause and prevention
- Every quarter, read the oldest entries and ask: "how did this turn out? Was the reasoning right?"
This is the minimum viable episodic memory layer. It is far less than Koylan's system but contains the core design move.
Key Points
- Episodic memory = judgment layer distinct from factual/semantic PKM content
- Concept from Tulving (1972); applied to PKM architecture by Koylan (2026)
- Three canonical logs: experiences (with emotional weight), decisions (with reasoning + outcome), failures (with root cause + prevention)
- Matters because decisions are the high-leverage data agents should reason from
- Complements rather than replaces semantic PKM (evergreen notes, reference material)
- Append-only, retroactive outcome tracking, cross-referenced to frameworks
- Privacy surface and habit sustainability are the main practical risks
Open Questions
- What is the minimum viable episodic memory practice that actually sustains and produces value?
- Can AI agents reliably extract episodic material from daily notes, or does it need explicit logging?
- How should episodic-memory logs interact with periodic reviews and retrospectives?
References
- Muratcan Koylan, "The File System Is the New Database" (2026-02)
- Endel Tulving, "Episodic and Semantic Memory" (1972) — the psychological distinction this borrows from
- Related decision-journal practices (Jim Collins, Daniel Kahneman on decision journals)