Topic
Persistent memory layers that let AI agents retain lessons, fixes and environment facts across sessions, forming a category alongside Mem0, Letta and LangMem.
Anthropic's dreams job reads an agent's memory store and 1 to 100 past sessions, then writes a reorganized copy. The original is never modified, so a team can review the copy and discard it.
Reality
Memory handoffs cut replaced agent context by 98.65% across 10,241 completed runs, according to an audit published on dev.to. For agent builders, the open risk now sits in whether the agent recalls the right stored text, since lossy summaries are no longer the failure point.
Reality
Throughline's on-call agent returns a three-way coverage verdict with every memory recall, so a timed-out search cannot pass as "no prior incidents". Its author hit the same error-as-empty trap in CockroachDB's managed MCP server, where failures come back as HTTP 200.
Reality
Every 10 minutes an unattended job summarizes Claude Code and Codex logs into an Obsidian vault and pushes the commit, so the pipeline treats the summarizer's own JSON as text that may carry pasted keys or injected instructions.
Reality
A runnable two-agent example moves every refusal into the catalog that vends credentials, because a key covers a storage prefix and there are no rows at the point of enforcement. The author works on that catalog.
Reality
A dev.to writeup by nasiko_labs describes a semantic memory tier whose first version inserted a row per session, so one agent could read "cautious" and "aggressive" at once. The rewrite keys each belief by fact type and archives the superseded version.
Reality
A Strands hook recorder writes each agent step into Neo4j through Neo4j Labs' agent-memory SDK. The audit query that finds every decision touching a bad source depends on a dict lookup keyed by tool name.
Reality
Both terminal agents recalled planted facts inside a single project across three two-session tests on identical Node repos. Only Grok Build carried a stated convention into an unrelated repo, at about a third of the reported cost.
Reality
Dhravya Shah spent three days probing the personal agent from the outside and reports git-tracked Markdown searched by keyword, with a background pass he clocked taking 23 hours and 16 minutes to commit one preference.
Reality
A dev.to walkthrough of Claude Code compaction says the system prompt and the last 10 to 15 turns survive while the messages between them are summarized and deleted without an error. Constraints you need later belong in a file.
Reality
An arXiv paper names the failure behavioral state decay and runs a side-car memory agent next to an unmodified action agent, reporting gains of 8.3 and 6.8 points of pass@1 on two long-horizon benchmarks.
Reality
In a dev.to account of self-reinforcing memory loops, an agent's own hedged inference is captured as a flat fact, retrieved a week later as context, and generalised until the store recommends replacing the cache layer.
Reality
Ramsauer and colleagues proved the equality in 2020, three years after attention was published. It holds for a single retrieval step inside a forward pass, and the capacity figures quoted alongside it come from a different model.
Reality
Mem0, Zep, Cognee, Letta, Supermemory and Mnemoverse all ship a knowledge graph, and each one's docs mean something different by a node. Read-time behaviour is the part a buyer can settle from the documentation.
Reality
The opencode-agent-memory plugin gives an OpenCode agent scoped Markdown blocks it rewrites through three tools, plus an opt-in journal it can search locally but never revise. The maintainer documents the failure modes himself.
Reality
Turing Post talked through agent feedback with Salesforce's chief AI scientist at Dreamforce. A correction lands in one of four places: the context, persistent memory, the software around the model, or the weights.
Reality
IBM Research's ALTK-Evolve distils an agent's own trajectories into scored guidelines and injects the top five at inference time, and a companion post puts a number on the reliability an average success rate hides.
Reality
Serge Kernbach's dev.to build notes put a 9B to 27B assistant on two RTX 4070s totalling 24 GB and argue that integration beats raw model quality. The figures he publishes are power draw and PCIe bandwidth.
Reality
Vendor pages for Claude Code, Cursor, Codex, VS Code and Windsurf, re-checked on 2026-09-12, show that all five resume a previous session. The complaint about agents forgetting points at the other three questions.
Reality
A developer instrumented 89 coding sessions to test whether rules the agent had already read changed what it did. Enforcement only started working once the rule was rewritten into something a hook could see.
Reality