Agent Memory
Agent memory is retained information that can become available to later work. The research problem is not just where information is stored: it is how information is admitted, selected, revised and consumed under a limited context budget.
Start with the agent-memory landscape, a source-grounded snapshot of 191 analyses from October 2026. Use the comparison table to find candidate mechanisms, then read the linked per-system records before making an architectural claim.
Focused comparisons
- How memory reaches the next decision — compare whole-file loading, requested retrieval, automatic injection and maintenance-model inputs in Hermes, A-MEM, Hindsight and Basic Memory.
- How corrections propagate through retained memory — compare temporal invalidation, dependent-record cleanup and explicit replacement in Graphiti, Hindsight and Mem0, then test the consumers that can still receive stale information.
Research questions
What information is retained?
Distinguish raw conversations, extracted facts, summaries, procedures and derived representations. A storage format does not by itself explain what later work can use. The memory accounts trace retained objects and consumers; the landscape's curation discussion compares the transformations identified in its frozen population.
Who decides what enters memory?
An automatic write can still depend on model judgment, fixed rules or caller selection. The write-admission comparison separates code-wired routes from instructions that merely ask an agent to save something. Follow each table row to its scope and evidence basis.
How is retained information selected for later use?
Ask who initiates a read, which selector runs, and who receives the result. Memory selected for a maintenance model is not necessarily delivered to an answering agent. Read How memory reaches the next decision for concrete contrasts, or begin with the landscape's read-back analysis and use the per-system accounts to trace the full route.
What happens when information changes or becomes invalid?
Updating a source, invalidating a fact and repairing every dependent summary are different operations. How corrections propagate traces those boundaries and proposes multi-interface experiments. The curation findings provide candidates for investigation, not a guarantee of complete correction. Check which consumer and derived representation each inspected route reaches.
How does experience modify future behavior?
Retaining a trace does not establish that later behavior changes. The trace-learning discussion distinguishes wired update routes and their limits. Follow the Workflow & Recovery topic when the experience comes from a multi-step run, and Knowledge & Verification when a retained conclusion must be checked before reliance.
What evidence establishes that memory improves performance?
A retrieval route, an observed read and a controlled improvement are separate findings. The frozen landscape reports static wiring or weaker evidence, not causal improvement. Experiments need to measure downstream decisions and compare conditions with and without the relevant memory route.
Find systems and data
- Current memory accounts and profiles — generated at each build from accepted analyses.
- Memory comparison research and datasets — landscape, table and CSV, with their snapshot boundaries.
- Systems Directory — whole-system overviews and frozen evidence boundaries.
- Methods & Data — evidence classifications, population selection and interpretation.