--- title: "Foundations" canonical: "https://current.tinyfish.ai/issue/25/foundations" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/25/foundations" issue: 25 section: "foundations" description: "Conceptual clarity earned from building at scale" authors: ["Nora Kaplan","Rina Takahashi"] published: 2026-04-19 updated: 2026-04-16 articles: 17 --- # TinyFish > CURRENT is a journal about AI agents in production, for the engineering and product leaders who have to make them work. Drawing on experience building enterprise web agents, it explains why the demo succeeded and the deployment didn't, and what that means for choosing vendors, briefing the board, and avoiding mistakes others have already made. [Read on the website]() ## Issue 25 — Foundations Conceptual clarity earned from building at scale ### The Unnamed Problem - [The Missing Half of Agent Memory]() — By [Nora Kaplan]() · Published 2026-04-15 · Format: The Unnamed Problem Every major agent memory effort shipped in the past year solves the same problem: retention. Longer context windows, durable checkpointing, consolidation pipelines that score and promote facts into permanent storage. The engineering is impressive, and it all faces one direction. ### The Concrete Case - [OpenClaw's Diary Timeline and the Line Between Transparency and Control]() — By [Rina Takahashi]() · Published 2026-04-15 · Format: The Concrete Case OpenClaw's April 9 "Dreaming" update shipped a UI called the Diary Timeline. Browse it and you'll find daily notes sitting alongside distilled versions of themselves. The distillation happened offline, through a consolidation pipeline the team modeled on biological sleep cycles. You're looking at the output of decisions about which experiences deserve to persist and which ones fade. ### Further Reading - **Agentic AI Evolution and the Security Claw** — [Source]() — Published 2026-04-15 Beck locates the risk not at agent failure but at delegation, where your identity becomes the agent's operating credential. - **Introducing the Agent Governance Toolkit** — [Source]() — Published 2026-04-15 Maps every OWASP agentic risk to a named countermeasure, making the taxonomy legible and exposing what still has no countermeasure at all. - **Stanford's 2026 AI Index in Charts** — [Source]() — Published 2026-04-15 The standout finding: benchmarks for agents and robots barely exist yet, which means governance frameworks are calibrating against fog. - **Governing Evolving Memory in LLM Agents** — [Source]() — Published 2026-04-15 Treats memory drift as its own governance surface, distinct from prompt injection. Most tooling hasn't even named this problem. - **Memory Scaling for AI Agents** — [Source]() — Published 2026-04-15 Production evidence that memory-equipped agents compound their own mistakes over time, with useful detail on what remediation actually costs. ### The Retention Bias Extraordinary energy is going into helping agents remember. Context windows have cleared a million tokens. Memory frameworks treat persistence as a selling point. Multi-agent architectures are multiplying memory state faster than anyone predicted. The other side of that coin is conspicuously bare. No standards body has convened a working group on forgetting. No cross-framework spec defines who controls what gets dropped, or how an operator would audit the decision. Five data points, placed side by side. They don't argue a thesis. The asymmetry does that on its own. - **Million-Token Context Windows Are Now Table Stakes** — [Source]() — Published 2026-04-15 Tag: Retention Race Five major models sit at 1M tokens. Meta's Scout leads at 10M. LangGraph ships durable execution with persistent memory as a headline feature, pulling tens of millions of monthly downloads. Retention has benchmarks, research literature, and a rapidly expanding tooling ecosystem behind it. - **Multi-Agent Workflows Grew 327% in Four Months** — [Source]() — Published 2026-04-15 Tag: Memory Growth Databricks tracked the surge across 20,000+ customers between June and October 2025. Every supervisor agent carries its own memory state alongside its sub-agents' states. Memory surface area now compounds with system complexity, not just agent count. - **OpenClaw Optimizes Remembering, Not Governing What's Forgotten** — [Source]() — Published 2026-04-15 Tag: Retention Optimization Its three-phase memory consolidation scores every candidate across six weighted signals. Entries below threshold expire. But expiration is a retention-tuning parameter, not a compliance mechanism. There are no deletion criteria, no forgetting spec, no audit trail for what disappears. - **No Standards Body Runs a Forgetting Working Group** — [Source]() — Published 2026-04-15 Tag: Standards Absence NIST's AI Agent Standards Initiative covers identity and security. NCCoE covers authorization. OWASP's agentic top ten treats memory poisoning as an attack vector. None of them addresses memory decay governance, deletion criteria, or forgetting authority. The topic has no institutional home. - **Who Decides What an Agent Forgets? Nobody** — [Source]() — Published 2026-04-15 Tag: Decay Ownership Every framework rolls its own decay heuristics. No cross-platform spec exists for memory deletion. No export standard. No external audit mechanism for what gets dropped or why. EU AI Act obligations arrive August 2026. Shared tooling to meet them does not. ### The Governance Surface Agent memory is getting more sophisticated every month. The governance around it is barely moving. We've been tracking five specific dimensions where agent memory operates without meaningful oversight. These aren't speculative concerns. You can trace each one through official documentation, published audits, and protocol roadmaps that exist right now: the scoring logic that decides what an agent keeps, the total absence of standards around what it drops, the permissions that never update when an agent's knowledge quietly changes underneath them. These gaps don't share an owner, a timeline, or a fix. Some are engineering problems. Some lack the vocabulary to even become problems yet. Here's where things stand. - **The Model Decides What Your Agent Remembers** — [Source]() — Published 2026-04-15 Tag: Scoring Criteria OpenClaw's April 2026 Dreaming update lays out a three-phase consolidation pipeline: six weighted signals, three hard promotion gates, all determining what reaches long-term memory. Per their docs, "phase policy, thresholds, and storage behavior are internal implementation details (not user-facing config)." Operators can preview outputs and roll them back. They cannot touch the scoring logic. - **OWASP Named It. Nobody Has Fixed It.** — [Source]() — Published 2026-04-15 Tag: Memory Poisoning OWASP's Top 10 for Agentic Applications lists Memory & Context Poisoning as ASI06. Research cited in the report shows over 95% injection success rates against production agents. The OWASP Agent Memory Guard project is building the first reference implementation, but a stable v1.0 isn't expected until Q4 2026. - **What Agents Forget Leaves No Trace Behind** — [Source]() — Published 2026-04-15 Tag: Audit Trails The MCP 2026 roadmap flags audit trails as an undefined enterprise gap for tool calls. Memory makes this worse: entries that fail consolidation promotion gates simply vanish from any external-facing surface. No Enterprise Working Group exists to address either layer yet. - **Permissions Stay Frozen While Memory Rewrites Itself** — [Source]() — Published 2026-04-15 Tag: Authorization Drift Agent permissions get scoped at session start or tool invocation. Never at memory revision time. So when consolidation reshapes what an agent durably knows, its authorization state just sits there unchanged. A Grantex Research audit of 30 open-source agent frameworks found 93% rely on unscoped API keys with no cascade revocation. - **Nobody Measures Whether Agent Memories Are True** — [Source]() — Published 2026-04-15 Tag: Fidelity Measurement Stanford's 2026 AI Index puts it plainly: "for complex, interactive technologies such as AI agents and robots, benchmarks barely exist yet." Memory fidelity, whether consolidated memories accurately reflect actual interaction history, isn't a named metric in any framework. Databricks has documented agents confidently citing their own prior errors as evidence. ## Other sections in Issue 25 - [Daily Brief]() - [Market Pulse]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()