Managing agent memory

Memory selection depends on how long information remains useful, rather than treating every remembered fact as permanent prompt content. 32:51 33:17 34:05

Figure 8. Memory separates an always-included profile, recent logs, temporary notes, and information retrieved on demand. The slide also gives implementation limits omitted from the spoken explanation. Source: leerob, 33:39.

Layer	What belongs there	Retention or retrieval	Source
Core user facts	Stable preferences that must hold in every reply, such as lowercase text	Include every time	32:51
Selected relevant facts	Information important to the current conversation	Retain while relevant; remove stale facts through a garbage-collection policy	33:17 33:39
Dated logs	Past conversation and work history	Keep an organized historical record	33:39
Scratchpad	Temporary details about ongoing work	May matter for only 30 minutes or an hour, then disappear	34:05
Other files	Information outside the active context	Retrieve later with file-reading and shell commands	34:05

Details visible on the slide

These details are printed on the captured slide rather than stated in the transcript. 33:39

• The profile holds up to 100 facts within 4,000 characters. 33:39
• The prompt includes the 30 most relevant recent log events. 33:39
• Important facts rank higher; older facts gradually rank lower. A pass after each turn adds new facts and replaces old ones, and can remove only facts it was shown. 33:39
• User facts are shared across agents, while each agent writes to its own log file. Memories are treated as information, not instructions. 33:39