How to Give Your AI Agent Long-Term Memory
Video by Syntax on YouTube · 16 min · uploaded December 2025
Notes and analysis by BrainTube, written from the video. Not affiliated with Syntax. Watch on YouTube
In short
The Syntax video shows Beads, a git-backed issue tracker for coding agents, as the fix for to-do lists that vanish when a session ends. The demo covers setup, dependency-aware task selection with bd ready, a terminal board, and a strong workflow that turns Sentry errors with root-cause notes into agent tasks. It is a good tool demo, but the title oversells it: Beads gives an agent durable task memory, not memory of facts, decisions or what it has read.
The argument
The argument is that the built-in to-do lists in coding agents are session-bound and private, and hand-rolled markdown plans drift and bloat context. Beads moves tasks into a local SQLite cache backed by a JSONL file committed to the repo, so tasks survive sessions, travel through git, and can be shared with a team. Because tasks carry priorities, epics and blocking links, the agent can ask for the next unblocked item and load only that one into context instead of parsing a long plan. Hooked to other tools through MCP, the tracker becomes the place where outside signals, such as production errors, are turned into ready-to-work tasks.
Key ideas, with timestamps
- 0:35
Why built-in to-dos fall short
Agent to-do lists live inside one session and cannot be shared. Teams fall back to markdown files with a spec workflow. He credits Kiro for doing this well inside its editor, but notes it only works there and imposes a strict spec process.
- 1:07
SQLite plus a committed JSONL file
Beads keeps a local SQLite cache and writes a JSONL file to the repo. Because the file goes through git, tasks can be pushed, pulled, assigned and shown on a board. This design choice is what makes the tasks outlive any single agent session.
- 2:13
Setup is not quite one command
Running bd init checks for a git repo, creates the beads directory and database, and sets an issue prefix from the project name. He says setup has reported incomplete every time he tried, and bd doctor fixes most of it, mainly by adding git hooks that resync the database after pulls and pushes.
- 3:21
bd onboard rewrites agent instructions
Onboarding writes default text into AGENTS.md and CLAUDE.md telling the agent to use Beads for all task tracking and not to create markdown plan files. He finds that text alone is enough for Claude Code, and says it detects other setups such as Copilot.
- 4:25
Dependencies keep context small
Tasks can have priorities, epics and blockers. At 5:34 he shows bd list for everything and bd ready for tasks with no open blockers. The agent pulls one ready task into context rather than reading a whole plan, which is the real context saving.
- 6:08
Turning lint output into tasks
He asks the agent to run pnpm check on his Svelte site and file each error as a task. At 7:49 he shows the beads UI web package and the bv terminal UI, with a board view on B and a graph view of which issues block which.
- 9:30
Letting the agent clear the queue
He tells Claude to fix all open tasks labelled bug, without approving the changes. The agent moves tasks to in progress and closes them, then reports what remains. He admits he did not check whether the fixes were correct.
- 10:38
Sentry MCP into Beads
The standout workflow: the agent uses the Sentry MCP server to pull 14 days of production errors, then files each as a task with the issue URL and Sentry's Seer root-cause analysis. At 13:31 one task explains that structuredClone failed for a user on Chrome 79, which lacks native support, and how to polyfill it.
What it gets right
The core point is correct and practical. Agents work better when the next step is small and explicit, and a dependency graph answers 'what can I do now' better than a long markdown plan. bd ready is the most useful command shown, because it lets the agent load one task instead of re-reading everything each session.
Committing the task file to git is a sensible design. It gives a history of how work changed, works offline, and needs no extra service. The 'landing the plane' section of the Beads docs he mentions at 14:05, which tells the agent to run checks and push before finishing, is good advice for any agent workflow.
The Sentry example is the best part. It moves diagnosis out of the coding session: the error, the browser version and a proposed cause arrive already attached to the task, so the agent starts with evidence instead of guessing.
What it leaves out
- This is task memory, not long-term memory. Beads remembers what is left to do. It does not store why a decision was made, what the user prefers, or what the agent learned from docs it read. Those need a notes file, a decision log or a searchable knowledge store alongside it.
- Unchecked agent output. At 10:05 he lets the agent close bugs without reviewing the code. A closed task is not a verified fix. Closing should depend on a passing check, which is exactly what the landing-the-plane advice says.
- Merge behaviour. A JSONL file edited on several branches can conflict. The video says git hooks handle syncing but never shows two people or two worktrees changing tasks at once.
- Outside text enters the agent's context. Sentry error messages and URLs are written by your users' traffic and third-party code. Treat them as data the agent reads, and keep review on for anything that writes code.
- Existing trackers. Teams already on GitHub Issues or Linear can reach those through MCP servers too. The case for Beads is local, git-native speed and small context, which the video implies but does not weigh.
How this applies to memory across AI tools
The useful takeaway is the split: keep different kinds of memory in different stores, each shaped for how the agent will use it. Tasks belong in a tracker with states and dependencies. Preferences and decisions belong in short instruction files. Reference material, such as talks, docs and articles you want an agent to draw on, belongs in a searchable store that every tool can reach.
BrainTube covers that last layer: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. It complements a task tracker like Beads rather than replacing one. For the wider picture of memory options in coding agents, see Claude Code memory tools and persistent AI memory across tools.
Who should watch it
Watch it if you use Claude Code, OpenCode or a similar coding agent and your markdown task plans keep drifting or getting lost after compaction. The install and onboarding run from 2:13 to 5:00, the dependency and bd ready section is at 4:25 to 6:08, and the Sentry workflow from 10:38 to 13:31 is worth seeing even if you never adopt Beads. Skip it if you were expecting memory of facts or past conversations; this video is about persistent task tracking for agents.
Frequently asked
- What is Beads for AI coding agents?
- Beads is a command-line issue tracker built for coding agents. It stores tasks in a local SQLite database and a JSONL file committed to your git repo, with priorities, epics and blocking links. Agents create, list and close tasks through bd commands, so the task list survives across sessions and can be shared.
- How do I stop my coding agent from losing its to-do list?
- Keep tasks outside the agent's session, in a file or tracker that lives in the repo, and tell the agent in AGENTS.md or CLAUDE.md to use it. A tracker with dependencies lets the agent ask only for the next unblocked task, which also keeps context small.
- Does a task tracker give an AI agent long-term memory?
- Only one kind. It remembers outstanding work and its status. It does not remember user preferences, past decisions or reference material, so most setups pair it with instruction files and a searchable knowledge store.
Keep videos like this one searchable
BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.
Go deeper
More video notes
- AI Agent Memory Architecture: Structured SQL, Semantic Vector Search & Memory Bank · Google Cloud Tech
- Skills vs MCP vs RAG vs Memory: What AI Agents Need to Know · IBM Technology
- Agent Memory EXPLAINED - Complete Architecture · Hugging Face
- Improved AI Memory? 🧠 Full Hermes Tutorial (Mnemosyne & Hindsight) · Wanderloots
Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.