Memory Scheduling
1. Capability Introduction
In MemOS, Memory Scheduling improves efficiency and accuracy by dynamically coordinating memories of different usage efficiencies (Parameter > Activated > Working > Other Plaintext). During conversations and tasks, it predicts which memories will be needed and preloads high-efficiency types like activated and working memories, accelerating the reasoning chain.
In complex interactions, if the system only relies on simple global search each time, it may:
- Be too slow: Waiting until after the user finishes asking to search causes high first-token latency.
- Be inaccurate: Too much history overwhelms key information, making retrieval difficult.
The role of scheduling is to give the system “instant readiness and fast response” capabilities:
- Preloading: At the start of the conversation, load the user’s commonly used background.
- Predictive invocation: Before the user finishes typing, prepare the memories that may be needed.
| Dimension | Explanation |
|---|---|
| What to schedule? | Parameter memory (long-term knowledge and skills) Activated memory (runtime KV cache and hidden states) Plaintext memory (externally editable facts, user preferences, retrieval snippets) Supports dynamic migration among Plaintext ⇆ Activated ⇆ Parameter; frequently used plaintext snippets can be compiled into KV cache in advance; stable templates can be deposited into parameters. |
| When to schedule? | When context and efficient memory are insufficient to support answering user queries, memory structures are optimized. Prepare memory content in advance according to the user’s intent and needs. During continuous queries, scheduling ensures high efficiency and accuracy in conversation scenarios. |
| Who to schedule for? | Current user, specific role agent, or shared cross-task context |
| What form to schedule into? | Memories are tagged with indicators such as heat, timeliness, and importance. The scheduler decides which to load first, which to cool down, and which to archive. |
When using MemOS cloud services, the role of scheduling can be observed through the performance of the searchMemory API:
- It quickly returns relevant memories, avoiding context breaks.
- Returned content is already optimized by the scheduler, ensuring results are relevant without overloading model input.
2. Example: Memory Scheduling in a Household Assistant Scenario
- Some time ago: the user was busy buying a house
User often said:
- “Check the average second-hand housing price in XX community.”
- “Remind me to view the house on Saturday.”
- “Record the latest change in mortgage rates.”
MemOS System Operation
- Initially, the system generated these items as Plaintext Memories.
- Since house-buying information was frequently mentioned, the scheduler judged it as a core theme and migrated these plaintexts into Activated Memories, making subsequent queries faster and more direct.
- Recently: the user bought the house and started decorating
User frequently mentioned:
- “Going to check tiles this weekend.”
- “Remind me to confirm water and electricity work with the renovation company.”
- “Record next week’s furniture delivery time.”
MemOS System Operation
- The system continued to generate new Plaintext Memories.
- The scheduler detected that “renovation” had become the new high-frequency theme and migrated these entries into Activated Memories.
- Meanwhile, previous “house-buying” activated memories were no longer frequently used and were automatically downgraded back to plaintext, reducing active memory usage.
- At the current moment: the user casually says—“I feel like too many things are piling up, help me organize them.”
Without scheduling, the system can only do a full-library retrieval, pulling out all possibly relevant memories:
- Checking tiles (renovation)
- Confirming water/electricity work (renovation)
- Furniture delivery (renovation)
- Checking housing prices (house-buying, outdated)
- Viewing houses (house-buying, outdated)
- Grocery shopping (daily life)
- Watching a movie (daily life)
With scheduling, the system can return more quickly
- Checking tiles
- Confirming water/electricity work
- Furniture delivery
👉 User experience UP
- Faster response (no need for full-library search).
- The listed items are exactly what they care about most → makes the assistant feel “very understanding.”
3. Advanced: Deep Customization
Developers can extend scheduling strategies to customize system behavior, mainly including:
| Extension Point | Configurable Content | Example Scenario |
|---|---|---|
| Scheduling Strategy | Define memory selection logic for different tasks | Conversation systems prioritize activated memories; research systems prioritize retrieving the latest plaintexts |
| Transformation Rules | Set conditions for cross-type migration | High-frequency FAQs → KV cache; stable paradigms → parameter modules |
| Context Binding | Bind memories to roles/users | Student users automatically load learning preferences; enterprise users load project archives |
| Permissions & Governance | Combine scheduling with access control and compliance checks | Medical records visible only to doctors; sensitive content not shareable across domains |
| Scheduling Metrics | Optimize scheduling based on access frequency and latency needs | High-frequency hot memories prioritized; low-frequency cold memories downgraded to archive |
4. Next Steps
Learn more about MemOS core capabilities:
5. Contact Us
Memory Production
The Memory Production module transforms raw messages or events into storable and retrievable memory units, serving as the starting point of the entire MemOS workflow.
Memory Recall and Instruction Completion
In MemOS, memory is not just about archiving information, but also about being dynamically retrieved when needed and transformed into executable input. This process is achieved through two closely connected steps: Memory Recall and Instruction Completion.